feat: CCCL CachingDeviceAllocator preload — 完整依赖链 288 files

从 cccl_upstream 递归追踪 cub/util_allocator.cuh 的全部 include 依赖:
  cub/         9 files (config, util_*, version, detect_cuda_runtime)
  cuda/        libcudacxx type_traits, concepts, algorithm, iterator...
  nv/          target macros, preprocessor

总计 288 个头文件 (1.4MB)，打包到 include/ 目录，编译时 -I include
即可完全脱离 CCCL 原始目录结构。

.cu 文件直接 #include <cub/util_allocator.cuh>，
走原版 CUB CachingDeviceAllocator，零 mock。

BI-V100 参数: growth=2 bins=[8..32] max_cached=8GB/device

qwen3_6_scripts/cccl_preload/include/cub/version.cuh

@@ -0,0 +1,65 @@
+// SPDX-FileCopyrightText: Copyright (c) 2011-2022, NVIDIA CORPORATION. All rights reserved.
+// SPDX-License-Identifier: BSD-3
+
+/*! \file version.cuh
+ *  \brief Compile-time macros encoding CUB release version
+ *
+ *         <cub/version.h> is the only CUB header that is guaranteed to
+ *         change with every CUB release.
+ *
+ */
+
+#pragma once
+
+// For _CCCL_IMPLICIT_SYSTEM_HEADER
+#include <cuda/__cccl_config>
+
+#if defined(_CCCL_IMPLICIT_SYSTEM_HEADER_GCC)
+#  pragma GCC system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_CLANG)
+#  pragma clang system_header
+#elif defined(_CCCL_IMPLICIT_SYSTEM_HEADER_MSVC)
+#  pragma system_header
+#endif // no system header
+
+#include <cuda/version>
+
+/*! \def CUB_VERSION
+ *  \brief The preprocessor macro \p CUB_VERSION encodes the version
+ *         number of the CUB library as MMMmmmpp.
+ *
+ *  \note CUB_VERSION is formatted as `MMMmmmpp`, which differs from `CCCL_VERSION` that uses `MMMmmmppp`.
+ *
+ *         <tt>CUB_VERSION % 100</tt> is the sub-minor version.
+ *         <tt>CUB_VERSION / 100 % 1000</tt> is the minor version.
+ *         <tt>CUB_VERSION / 100000</tt> is the major version.
+ */
+#define CUB_VERSION 300500 // macro expansion with ## requires this to be a single value
+
+/*! \def CUB_MAJOR_VERSION
+ *  \brief The preprocessor macro \p CUB_MAJOR_VERSION encodes the
+ *         major version number of the CUB library.
+ */
+#define CUB_MAJOR_VERSION (CUB_VERSION / 100000)
+
+/*! \def CUB_MINOR_VERSION
+ *  \brief The preprocessor macro \p CUB_MINOR_VERSION encodes the
+ *         minor version number of the CUB library.
+ */
+#define CUB_MINOR_VERSION (CUB_VERSION / 100 % 1000)
+
+/*! \def CUB_SUBMINOR_VERSION
+ *  \brief The preprocessor macro \p CUB_SUBMINOR_VERSION encodes the
+ *         sub-minor version number of the CUB library.
+ */
+#define CUB_SUBMINOR_VERSION (CUB_VERSION % 100)
+
+/*! \def CUB_PATCH_NUMBER
+ *  \brief The preprocessor macro \p CUB_PATCH_NUMBER encodes the
+ *         patch number of the CUB library.
+ */
+#define CUB_PATCH_NUMBER 0
+
+static_assert(CUB_MAJOR_VERSION == CCCL_MAJOR_VERSION);
+static_assert(CUB_MINOR_VERSION == CCCL_MINOR_VERSION);
+static_assert(CUB_SUBMINOR_VERSION == CCCL_PATCH_VERSION);