[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/benchmarks/bench/scan/exclusive/base.cuh

@@ -0,0 +1,88 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2026, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+#pragma once
+
+#include <cub/device/device_scan.cuh>
+
+#include <cuda/std/__functional/invoke.h>
+
+#include <nvbench_helper.cuh>
+
+#include "../policy_selector.h"
+
+template <typename T, typename OffsetT>
+static void basic(nvbench::state& state, nvbench::type_list<T, OffsetT>)
+try
+{
+  using init_value_t             = T;
+  using accum_t [[maybe_unused]] = ::cuda::std::__accumulator_t<op_t, init_value_t, T>;
+  using offset_t                 = cub::detail::choose_offset_t<OffsetT>;
+#if USES_LOOKAHEAD()
+  static_assert(sizeof(offset_t) == sizeof(size_t)); // lookahead scan uses size_t internally
+#endif // USES_LOOKAHEAD()
+
+  const auto elements = static_cast<std::size_t>(state.get_int64("Elements{io}"));
+  if (sizeof(offset_t) == 4 && elements > std::numeric_limits<offset_t>::max())
+  {
+    state.skip("Skipping: input size exceeds 32-bit offset type capacity.");
+    return;
+  }
+
+  thrust::device_vector<T> input = generate(elements);
+  thrust::device_vector<T> output(elements);
+
+  const T* d_input = thrust::raw_pointer_cast(input.data());
+  T* d_output      = thrust::raw_pointer_cast(output.data());
+
+  state.add_element_count(elements);
+  state.add_global_memory_reads<T>(elements, "Size");
+  state.add_global_memory_writes<T>(elements);
+
+  caching_allocator_t alloc;
+  state.exec(nvbench::exec_tag::gpu | nvbench::exec_tag::no_batch, [&](nvbench::launch& launch) {
+    auto env = cub_bench_env(
+      alloc,
+      launch
+#if !TUNE_BASE
+      ,
+      cuda::execution::tune(policy_selector<accum_t>{})
+#endif // !TUNE_BASE
+    );
+    _CCCL_TRY_CUDA_API(
+      cub::DeviceScan::ExclusiveScan,
+      "ExclusiveScan failed",
+      d_input,
+      d_output,
+      op_t{},
+      init_value_t{},
+      static_cast<offset_t>(input.size()),
+      env);
+  });
+}
+catch (const std::bad_alloc&)
+{
+  state.skip("Skipping: out of memory.");
+}
+
+// __half and __nv_bfloat16 are added for full (non-tuning) runs; CUB has fast paths for them (see #9587).
+#ifdef TUNE_T
+using value_types = nvbench::type_list<TUNE_T>;
+#else
+using value_types =
+  push_back_t<all_types
+#  if _CCCL_HAS_NVFP16() && _CCCL_CTK_AT_LEAST(12, 2)
+              ,
+              __half
+#  endif
+#  if _CCCL_HAS_NVBF16() && _CCCL_CTK_AT_LEAST(12, 2)
+              ,
+              __nv_bfloat16
+#  endif
+              >;
+#endif
+
+NVBENCH_BENCH_TYPES(basic, NVBENCH_TYPE_AXES(value_types, scan_offset_types))
+  .set_name("base")
+  .set_type_axes_names({"T{ct}", "OffsetT{ct}"})
+  .add_int64_power_of_two_axis("Elements{io}", nvbench::range(16, 32, 4));