[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/sum.cu

@@ -0,0 +1,22 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2023, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+// Tuning parameters found for signed integer types apply equally for unsigned integer types
+
+#include <nvbench_helper.cuh>
+
+// This benchmark tunes the old, non-lookahead scan implementation. Using it for benchmarking, will pick the lookahead
+// implementation on SM100+, but it's better to use the sum.lookahead.cu benchmark instead, which uses a single OffsetT.
+
+// %RANGE% TUNE_ITEMS ipt 7:24:1
+// %RANGE% TUNE_THREADS tpb 128:1024:32
+// %RANGE% TUNE_MAGIC_NS ns 0:2048:4
+// %RANGE% TUNE_DELAY_CONSTRUCTOR_ID dcid 0:7:1
+// %RANGE% TUNE_L2_WRITE_LATENCY_NS l2w 0:1200:5
+// %RANGE% TUNE_TRANSPOSE trp 0:1:1
+// %RANGE% TUNE_LOAD ld 0:1:1
+
+#define USES_LOOKAHEAD() 0
+using op_t              = ::cuda::std::plus<>;
+using scan_offset_types = offset_types;
+#include "base.cuh"