[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/libcudacxx/include/cuda/__runtime/api_wrapper.h

@@ -0,0 +1,60 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___RUNTIME_API_WRAPPER_H
+#define _CUDA___RUNTIME_API_WRAPPER_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/std/__exception/cuda_error.h>
+#include <cuda/std/__exception/exception_macros.h>
+
+#define _CCCL_TRY_CUDA_API(_NAME, _MSG, ...)                   \
+  do                                                           \
+  {                                                            \
+    const ::cudaError_t __status = _NAME(__VA_ARGS__);         \
+    if (__status != ::cudaSuccess)                             \
+    {                                                          \
+      ::cudaGetLastError(); /* clear CUDA error state */       \
+      _CCCL_THROW(::cuda::cuda_error, __status, _MSG, #_NAME); \
+    }                                                          \
+  } while (0)
+
+#define _CCCL_ASSERT_CUDA_API(_NAME, _MSG, ...)                         \
+  do                                                                    \
+  {                                                                     \
+    [[maybe_unused]] const ::cudaError_t __status = _NAME(__VA_ARGS__); \
+    ::cudaGetLastError(); /* clear CUDA error state */                  \
+    _CCCL_ASSERT(__status == cudaSuccess, _MSG);                        \
+  } while (0)
+
+#define _CCCL_LOG_CUDA_API(_NAME, _MSG, ...)                                       \
+  [&]() {                                                                          \
+    const ::cudaError_t __status = _NAME(__VA_ARGS__);                             \
+    if (__status != ::cudaSuccess)                                                 \
+    {                                                                              \
+      ::cuda::__msg_storage __msg_buffer;                                          \
+      ::cuda::__detail::__format_cuda_error(__msg_buffer, __status, _MSG, #_NAME); \
+      ::fprintf(stderr, "%s\n", __msg_buffer.__buffer);                            \
+      ::fflush(stderr);                                                            \
+    }                                                                              \
+    ::cudaGetLastError(); /* clear CUDA error state */                             \
+    return __status;                                                               \
+  }()
+
+#endif //_CUDA___RUNTIME_API_WRAPPER_H

cccl_upstream/libcudacxx/include/cuda/__runtime/ensure_current_context.h

@@ -0,0 +1,99 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___RUNTIME_ENSURE_CURRENT_CONTEXT_H
+#define _CUDA___RUNTIME_ENSURE_CURRENT_CONTEXT_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
+
+#  include <cuda/__device/device_ref.h>
+#  include <cuda/__device/physical_device.h>
+#  include <cuda/__driver/driver_api.h>
+
+#  include <cuda/std/__cccl/prologue.h>
+
+#  ifndef _CCCL_DOXYGEN_INVOKED // Do not document
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+class stream_ref;
+
+//! @brief RAII helper which on construction sets the current context to the specified one.
+//! It sets the state back on destruction.
+//!
+struct [[maybe_unused]] __ensure_current_context
+{
+  //! @brief Construct a new `__ensure_current_context` object and switch to the primary context of the specified
+  //!        device.
+  //!
+  //! @param new_device The device to switch the context to
+  //!
+  //! @throws cuda_error if the context switch fails
+  _CCCL_HOST_API explicit __ensure_current_context(device_ref __new_device)
+  {
+    auto __ctx = ::cuda::__physical_devices()[__new_device.get()].__primary_context();
+    ::cuda::__driver::__ctxPush(__ctx);
+  }
+
+  //! @brief Construct a new `__ensure_current_context` object and switch to the specified
+  //!        context.
+  //!
+  //! @param ctx The context to switch to
+  //!
+  //! @throws cuda_error if the context switch fails
+  _CCCL_HOST_API explicit __ensure_current_context(::CUcontext __ctx)
+  {
+    ::cuda::__driver::__ctxPush(__ctx);
+  }
+
+  //! @brief Construct a new `__ensure_current_context` object and switch to the context
+  //!        under which the specified stream was created.
+  //!
+  //! @param stream Stream indicating the context to switch to
+  //!
+  //! @throws cuda_error if the context switch fails
+  _CCCL_HOST_API explicit __ensure_current_context(stream_ref __stream);
+
+  __ensure_current_context(__ensure_current_context&&)                 = delete;
+  __ensure_current_context(__ensure_current_context const&)            = delete;
+  __ensure_current_context& operator=(__ensure_current_context&&)      = delete;
+  __ensure_current_context& operator=(__ensure_current_context const&) = delete;
+
+  //! @brief Destroy the `__ensure_current_context` object and switch back to the original
+  //!        context.
+  //!
+  //! @throws cuda_error if the device switch fails. If the destructor is called
+  //!         during stack unwinding, the program is automatically terminated.
+  _CCCL_HOST_API ~__ensure_current_context() noexcept(false)
+  {
+    // TODO would it make sense to assert here that we pushed and popped the same thing?
+    ::cuda::__driver::__ctxPop();
+  }
+};
+
+_CCCL_END_NAMESPACE_CUDA
+
+#  endif // _CCCL_DOXYGEN_INVOKED
+
+#  include <cuda/std/__cccl/epilogue.h>
+
+#endif // _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
+
+#endif // _CUDA___RUNTIME_ENSURE_CURRENT_CONTEXT_H

cccl_upstream/libcudacxx/include/cuda/__runtime/types.h

@@ -0,0 +1,47 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of libcu++, the C++ Standard Library for your entire system,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef _CUDA___RUNTIME_TYPES_H
+#define _CUDA___RUNTIME_TYPES_H
+
+#include <cuda/std/detail/__config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#if _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
+
+#  include <cuda/std/__cccl/prologue.h>
+
+_CCCL_BEGIN_NAMESPACE_CUDA
+
+_CCCL_DIAG_PUSH
+_CCCL_DIAG_SUPPRESS_CLANG("-Wmissing-braces")
+// clang complains about missing braces in CUmemLocation constructor but GCC complains if we add them
+
+using memory_location = ::cudaMemLocation;
+#  if _CCCL_CTK_AT_LEAST(12, 2)
+inline constexpr memory_location host_memory_location = {::cudaMemLocationTypeHost, 0};
+#  endif // _CCCL_CTK_AT_LEAST(12, 2)
+
+_CCCL_DIAG_POP
+
+_CCCL_END_NAMESPACE_CUDA
+
+#endif // _CCCL_HAS_CTK() && !_CCCL_COMPILER(NVRTC)
+
+#include <cuda/std/__cccl/epilogue.h>
+
+#endif // __CUDA___RUNTIME_TYPES_H