[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/adjacent_difference/subtract_left.cu

@@ -0,0 +1,71 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2023, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+#include <cub/device/device_adjacent_difference.cuh>
+
+#include <nvbench_helper.cuh>
+
+// %RANGE% TUNE_ITEMS_PER_THREAD ipt 7:24:1
+// %RANGE% TUNE_THREADS_PER_BLOCK tpb 128:1024:32
+
+#if !TUNE_BASE
+struct policy_selector_t
+{
+  [[nodiscard]] _CCCL_HOST_DEVICE constexpr auto operator()(cuda::compute_capability) const
+    -> cub::AdjacentDifferencePolicy
+  {
+    return {TUNE_THREADS_PER_BLOCK,
+            TUNE_ITEMS_PER_THREAD,
+            cub::BLOCK_LOAD_WARP_TRANSPOSE,
+            cub::LOAD_CA,
+            cub::BLOCK_STORE_WARP_TRANSPOSE};
+  }
+};
+#endif // !TUNE_BASE
+
+template <class T, class OffsetT>
+void left(nvbench::state& state, nvbench::type_list<T, OffsetT>)
+{
+  using input_it_t      = const T*;
+  using output_it_t     = T*;
+  using difference_op_t = ::cuda::std::minus<>;
+  using offset_t        = cub::detail::choose_offset_t<OffsetT>;
+
+  const auto elements         = static_cast<std::size_t>(state.get_int64("Elements{io}"));
+  thrust::device_vector<T> in = generate(elements);
+  thrust::device_vector<T> out(elements, thrust::no_init);
+
+  input_it_t d_in   = thrust::raw_pointer_cast(in.data());
+  output_it_t d_out = thrust::raw_pointer_cast(out.data());
+
+  state.add_element_count(elements);
+  state.add_global_memory_reads<T>(elements);
+  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_t{})
+#endif // !TUNE_BASE
+    );
+    _CCCL_TRY_CUDA_API(
+      cub::DeviceAdjacentDifference::SubtractLeftCopy,
+      "SubtractLeftCopy failed",
+      d_in,
+      d_out,
+      static_cast<offset_t>(elements),
+      difference_op_t{},
+      env);
+  });
+}
+
+using types = nvbench::type_list<int32_t>;
+
+NVBENCH_BENCH_TYPES(left, NVBENCH_TYPE_AXES(types, offset_types))
+  .set_name("base")
+  .set_type_axes_names({"T{ct}", "OffsetT{ct}"})
+  .add_int64_power_of_two_axis("Elements{io}", nvbench::range(16, 28, 4));