[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/cub/cub/detail/deferred_parameter.cuh

@@ -0,0 +1,82 @@
+// SPDX-FileCopyrightText: Copyright (c) 2026, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+
+#pragma once
+
+#include <cub/config.cuh>
+
+#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 <cub/detail/choose_offset.cuh>
+
+#include <cuda/__argument/argument.h>
+#include <cuda/std/__type_traits/is_same.h>
+#include <cuda/std/__utility/declval.h>
+
+CUB_NAMESPACE_BEGIN
+
+namespace detail
+{
+#if !_CCCL_COMPILER(NVRTC)
+// Preserve deferred problem sizes for dispatch and canonicalize immediate values to CUB's offset type.
+template <typename NumItemsT>
+[[nodiscard]] CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE auto make_num_items_dispatch_arg(NumItemsT num_items) noexcept
+{
+  using args_traits_t = ::cuda::args::__traits<NumItemsT>;
+
+  if constexpr (args_traits_t::is_deferred)
+  {
+    return num_items;
+  }
+  else
+  {
+    using offset_t = choose_offset_t<typename args_traits_t::element_type>;
+    return static_cast<offset_t>(::cuda::args::__unwrap(num_items));
+  }
+}
+
+// Forms a kernel parameter from a single-value argument without reading a deferred source.
+// Immediate values are converted to TargetT. Deferred arguments are unwrapped to their source, erasing bounds from
+// the kernel type and payload.
+template <typename TargetT, typename ParameterT>
+[[nodiscard]] CUB_RUNTIME_FUNCTION _CCCL_FORCEINLINE constexpr auto parameter_from_host(ParameterT parameter) noexcept
+{
+  using args_traits_t = ::cuda::args::__traits<ParameterT>;
+  static_assert(args_traits_t::is_single_value, "parameter must contain a single value");
+
+  if constexpr (args_traits_t::is_deferred)
+  {
+    return ::cuda::args::__unwrap(parameter);
+  }
+  else
+  {
+    return static_cast<TargetT>(::cuda::args::__unwrap(parameter));
+  }
+}
+
+template <typename TargetT, typename ParameterT>
+using parameter_from_host_t = decltype(parameter_from_host<TargetT>(::cuda::std::declval<ParameterT>()));
+#endif // !_CCCL_COMPILER(NVRTC)
+
+// Forms a value from a kernel parameter, reading element zero when the parameter is a deferred source.
+template <typename TargetT, typename ParameterT>
+[[nodiscard]] _CCCL_DEVICE_API _CCCL_FORCEINLINE TargetT parameter_from_device(ParameterT parameter) noexcept
+{
+  if constexpr (::cuda::std::is_same_v<ParameterT, TargetT>)
+  {
+    return parameter;
+  }
+  else
+  {
+    return static_cast<TargetT>(parameter[0]);
+  }
+}
+} // namespace detail
+
+CUB_NAMESPACE_END