[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/thrust/benchmarks/bench/unique/basic.cu

@@ -0,0 +1,40 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2023, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+#include <thrust/device_vector.h>
+#include <thrust/execution_policy.h>
+#include <thrust/unique.h>
+
+#include "nvbench_helper.cuh"
+
+template <typename T>
+static void basic(nvbench::state& state, nvbench::type_list<T>)
+{
+  const auto elements = static_cast<std::size_t>(state.get_int64("Elements"));
+
+  const std::size_t min_segment_size = 1;
+  const std::size_t max_segment_size = static_cast<std::size_t>(state.get_int64("MaxSegSize"));
+
+  thrust::device_vector<T> input = generate.uniform.key_segments(elements, min_segment_size, max_segment_size);
+  thrust::device_vector<T> output(elements);
+
+  caching_allocator_t alloc;
+  // not a warm-up run, we need to run once to determine the size of the output
+  const auto new_end             = thrust::unique_copy(policy(alloc), input.cbegin(), input.cend(), output.begin());
+  const std::size_t unique_items = ::cuda::std::distance(output.begin(), new_end);
+
+  state.add_element_count(elements);
+  state.add_global_memory_reads<T>(elements);
+  state.add_global_memory_writes<T>(unique_items);
+
+  state.exec(nvbench::exec_tag::gpu | nvbench::exec_tag::no_batch | nvbench::exec_tag::sync,
+             [&](nvbench::launch& launch) {
+               thrust::unique_copy(policy(alloc, launch), input.cbegin(), input.cend(), output.begin());
+             });
+}
+
+NVBENCH_BENCH_TYPES(basic, NVBENCH_TYPE_AXES(fundamental_types))
+  .set_name("base")
+  .set_type_axes_names({"T{ct}"})
+  .add_int64_power_of_two_axis("Elements", nvbench::range(16, 28, 4))
+  .add_int64_power_of_two_axis("MaxSegSize", {1, 4, 8});

cccl_upstream/thrust/benchmarks/bench/unique/by_key.cu

@@ -0,0 +1,58 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2023, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+#include <thrust/device_vector.h>
+#include <thrust/execution_policy.h>
+#include <thrust/unique.h>
+
+#include "nvbench_helper.cuh"
+
+template <class KeyT, class ValueT>
+static void basic(nvbench::state& state, nvbench::type_list<KeyT, ValueT>)
+{
+  const auto elements = static_cast<std::size_t>(state.get_int64("Elements"));
+
+  const std::size_t min_segment_size = 1;
+  const std::size_t max_segment_size = static_cast<std::size_t>(state.get_int64("MaxSegSize"));
+
+  thrust::device_vector<KeyT> in_keys = generate.uniform.key_segments(elements, min_segment_size, max_segment_size);
+  thrust::device_vector<KeyT> out_keys(elements);
+  thrust::device_vector<ValueT> in_vals(elements);
+  thrust::device_vector<ValueT> out_vals(elements);
+
+  caching_allocator_t alloc;
+  // not a warm-up run, we need to run once to determine the size of the output
+  const auto [new_key_end, new_val_end] = thrust::unique_by_key_copy(
+    policy(alloc), in_keys.cbegin(), in_keys.cend(), in_vals.cbegin(), out_keys.begin(), out_vals.begin());
+
+  const std::size_t unique_elements = ::cuda::std::distance(out_keys.begin(), new_key_end);
+  state.add_element_count(elements);
+  state.add_global_memory_reads<KeyT>(elements);
+  state.add_global_memory_writes<KeyT>(unique_elements);
+  state.add_global_memory_reads<ValueT>(elements);
+  state.add_global_memory_writes<ValueT>(unique_elements);
+
+  state.exec(
+    nvbench::exec_tag::gpu | nvbench::exec_tag::no_batch | nvbench::exec_tag::sync, [&](nvbench::launch& launch) {
+      thrust::unique_by_key_copy(
+        policy(alloc, launch), in_keys.cbegin(), in_keys.cend(), in_vals.cbegin(), out_keys.begin(), out_vals.begin());
+    });
+}
+
+using key_types =
+  nvbench::type_list<int8_t,
+                     int16_t,
+                     int32_t,
+                     int64_t
+#if _CCCL_HAS_INT128()
+                     ,
+                     int128_t
+#endif
+                     >;
+using value_types = all_types;
+
+NVBENCH_BENCH_TYPES(basic, NVBENCH_TYPE_AXES(key_types, value_types))
+  .set_name("base")
+  .set_type_axes_names({"KeyT{ct}", "ValueT{ct}"})
+  .add_int64_power_of_two_axis("Elements", nvbench::range(16, 28, 4))
+  .add_int64_power_of_two_axis("MaxSegSize", {1, 8});