[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/stf/reclaiming/graph.cu

@@ -0,0 +1,74 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDASTF 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) 2022-2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#include <cuda/experimental/__stf/graph/graph_ctx.cuh>
+
+#include <iostream>
+
+#if !_CCCL_COMPILER(MSVC)
+using namespace cuda::experimental::stf;
+
+__global__ void kernel()
+{
+  // No-op
+}
+#endif // !_CCCL_COMPILER(MSVC)
+
+int main([[maybe_unused]] int argc, [[maybe_unused]] char** argv)
+{
+// TODO fix setenv
+#if !_CCCL_COMPILER(MSVC)
+  int nblocks       = 4;
+  size_t block_size = 1024 * 1024;
+
+  if (argc > 1)
+  {
+    nblocks = atoi(argv[1]);
+  }
+
+  if (argc > 2)
+  {
+    block_size = atoi(argv[2]);
+  }
+
+  // At most 1 buffer is allocated at the same time
+  setenv("MAX_ALLOC_CNT", "1", 0);
+
+  graph_ctx ctx;
+
+  ::std::vector<logical_data<slice<char>>> handles(nblocks);
+
+  char* h_buffer = new char[nblocks * block_size];
+
+  for (int i = 0; i < nblocks; i++)
+  {
+    handles[i] = ctx.logical_data(make_slice(&h_buffer[i * block_size], block_size));
+    handles[i].set_symbol("D_" + std::to_string(i));
+  }
+
+  // We only 2 buffers, we are forced to reuse the buffer from D0 for D2
+  for (int i = 0; i < 3; i++)
+  {
+    ctx.task(handles[i % nblocks].rw())->*[&](cudaStream_t s, auto /*unused*/) {
+      kernel<<<1, 1, 0, s>>>();
+    };
+  }
+
+  ctx.submit();
+
+  if (argc > 3)
+  {
+    std::cout << "Generating DOT output in " << argv[3] << '\n';
+    ctx.print_to_dot(argv[1]);
+  }
+
+  ctx.finalize();
+#endif // !_CCCL_COMPILER(MSVC)
+}