fix: CUB_NS_QUALIFIER for wrapped namespace + drop _CCCL_COMPILER_GCC

CoreX clang++ 不是 GCC，-D_CCCL_COMPILER_GCC=1 和 CCCL 自己的
compiler detection 冲突。

CUB_WRAPPED_NAMESPACE=cccl_preload 使得命名空间变成 cccl_preload::cub，
用 CUB_NS_QUALIFIER 宏自动解析正确的命名空间。

qwen3_6_scripts/cccl_preload/build_cccl_preload.sh

@@ -72,8 +72,7 @@ COMMON_FLAGS=(
     # Suppress CCCL warnings that don't affect correctness
     -Wno-unused-function
     -Wno-unknown-pragmas
-    # CUB needs these for non-NVCC compilers
-    -D_CCCL_COMPILER_GCC=1
+    # Non-NVCC: disable arch magic, wrap namespace to avoid ODR conflicts
     -D__CUDA_ARCH_LIST__=700
     -DCUB_DISABLE_NAMESPACE_MAGIC
     -DCUB_WRAPPED_NAMESPACE=cccl_preload

qwen3_6_scripts/cccl_preload/cccl_allocator_preload.cu

@@ -41,8 +41,10 @@ static constexpr unsigned int ALLOC_MAX_BIN      = 32;  /* 4 GB */
 static constexpr size_t       ALLOC_MAX_CACHED   = (size_t)8 * 1024 * 1024 * 1024; /* 8GB */
 
 /* ---- Global allocator singleton ---- */
-static cub::CachingDeviceAllocator& get_allocator() {
-    static cub::CachingDeviceAllocator instance(
+using CubAllocator = CUB_NS_QUALIFIER::CachingDeviceAllocator;
+
+static CubAllocator& get_allocator() {
+    static CubAllocator instance(
         ALLOC_BIN_GROWTH,
         ALLOC_MIN_BIN,
         ALLOC_MAX_BIN,