[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/cudax/test/common/group_testing.cuh

@@ -0,0 +1,103 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// 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) 2026 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef COMMON_GROUP_CUH
+#define COMMON_GROUP_CUH
+
+#include <cuda/barrier>
+#include <cuda/std/cstddef>
+#include <cuda/std/type_traits>
+#include <cuda/warp>
+
+#include <cuda/experimental/group.cuh>
+
+#include "testing.cuh"
+
+namespace
+{
+template <class T, cuda::std::size_t Id>
+__device__ T global_barriers_storage;
+
+//! @brief Returns reference to an array of N cuda::barrier objects with suitable thread scope for level allocated in
+//!        suitable address space (shared or device memory). Id parameter can be used to create unique object.
+template <cuda::std::size_t N, cuda::std::size_t Id = 0, class Level>
+__device__ auto& get_barriers(const Level& level) noexcept
+{
+  constexpr auto scope = cudax::__minimum_required_scope_for<Level>();
+
+  using Barrier         = cuda::barrier<scope>;
+  using BarriersStorage = cuda::std::aligned_storage_t<N * sizeof(Barrier), alignof(Barrier)>;
+
+  if constexpr (scope >= cuda::thread_scope_block)
+  {
+    __shared__ BarriersStorage shared_barriers_storage;
+    return reinterpret_cast<Barrier(&)[N]>(shared_barriers_storage);
+  }
+  else
+  {
+    return reinterpret_cast<Barrier(&)[N]>(global_barriers_storage<BarriersStorage, Id>);
+  }
+}
+
+struct ThreadsInWarpMappingResult
+{
+  __device__ static constexpr ::cuda::std::size_t static_group_count()
+  {
+    return 1;
+  }
+
+  __device__ unsigned group_count() const
+  {
+    return 1;
+  }
+
+  __device__ unsigned group_rank() const
+  {
+    return 0;
+  }
+
+  __device__ static constexpr ::cuda::std::size_t static_unit_count()
+  {
+    return 32;
+  }
+
+  __device__ unsigned unit_count() const
+  {
+    return 32;
+  }
+
+  __device__ unsigned unit_rank() const
+  {
+    return cuda::gpu_thread.rank_as<unsigned>(cuda::warp);
+  }
+
+  __device__ cuda::device::lane_mask lane_mask() const noexcept
+  {
+    return cuda::device::lane_mask::all();
+  }
+
+  __device__ bool is_valid() const
+  {
+    return true;
+  }
+
+  __device__ static constexpr bool is_always_exhaustive() noexcept
+  {
+    return true;
+  }
+
+  __device__ static constexpr bool is_always_contiguous() noexcept
+  {
+    return true;
+  }
+};
+} // namespace
+
+#endif // COMMON_GROUP_CUH