Fix CUB compilation errors with CUDA 12.x by removing namespace wrapping

[ZhuYanzhen1]

Problem

When compiling DREAMPlace with CUDA 12.x, the build fails with numerous errors from CUB headers such as agent_three_way_partition.cuh and tuning_three_way_partition.cuh. Example errors include:

error: name followed by "::" must be a class or namespace name
  struct accumulator_pack_base_t<OffsetT, typename cuda::std::enable_if<sizeof(OffsetT) == 4>::type>
                                                         ^

Root Cause

The issue stems from two sources:

1. Namespace wrapping incompatibility: The project's utils_cub.cuh header wraps CUB includes with CUB_NS_PREFIX/CUB_NS_POSTFIX macros to place CUB inside the DreamPlace namespace. This technique was compatible with older CUB versions but breaks with CUDA 11+ where CUB uses the cuda::std namespace internally, which cannot be wrapped.

2. Potential version conflict: For CUDA 11+, the system CUB should be used exclusively, but the build configuration did not properly exclude the bundled old CUB (v1.8.0 from 2018) from include paths.

Solution

1. Modified dreamplace/ops/utility/src/utils_cub.cuh: Added conditional compilation to skip namespace wrapping for CUDA 11+. For CUDA 11+, we include the system CUB directly and create a namespace alias for compatibility. For CUDA 10 and below, we continue using the bundled CUB with namespace wrapping.

2. Modified CMakeLists.txt: Enhanced the CUB configuration to explicitly exclude the bundled CUB from CMAKE_PREFIX_PATH when using CUDA 11+, preventing accidental inclusion of the old CUB headers.