fix: use cccl_preload::cub namespace — CUB_WRAPPED_NAMESPACE requires it

CUB_DISABLE_NAMESPACE_MAGIC requires CUB_WRAPPED_NAMESPACE.
CUB_WRAPPED_NAMESPACE=cccl_preload wraps cub into cccl_preload::cub.
Source must use cccl_preload::cub::CachingDeviceAllocator.

qwen3_6_scripts/cccl_preload/build_cccl_preload.sh

@@ -75,6 +75,7 @@ COMMON_FLAGS=(
     # CUB needs this for non-NVCC compilers
     -D__CUDA_ARCH_LIST__=700
     -DCUB_DISABLE_NAMESPACE_MAGIC
+    -DCUB_WRAPPED_NAMESPACE=cccl_preload
 )
 
 if [[ "${CXX}" == *clang++* ]]; then