[BUG]: Fix MSVC build of CUB tests with RDC=true

[gevtushenko]

Is this a duplicate?

• I confirmed there appear to be no duplicate issues for this bug and that I agree to the Code of Conduct

Type of Bug

Compile-time Error

Component

CUB

Describe the bug

Recently, we've fixed an issue related to nvcc being unable to compile Catch2 main function. The fix consisted of separating Catch2 main function into a CXX target. This might've introduced a different issue. When building with MSVC / RDC=true, the following error occurs:

error LNK2019: unresolved external symbol __cudaRegisterLinkedBinary referenced in function 
"void __cdecl __nv_cudaEntityRegisterCallback(void * *)"

It looks like CXX linker is being selected. Apart from fixing this issue, we might want to separate Batch2 binaries into cub.cpp14.catch2_test.cdp_0 and cub.cpp14.catch2_test.cdp_1. This change should simplify tests development. Currently, it's easy to violate ODR when writing a CDP test. Developer should ensure that all of the test functions have internal visibility. Separating Catch2 binaries removes this requirement.

How to Reproduce

cmake -DCCCL_ENABLE_CUB=YES -DCCCL_ENABLE_THRUST=NO -DCCCL_ENABLE_LIBCUDACXX=NO -DCMAKE_CUDA_ARCHITECTURES=89 -DCUB_ENABLE_RDC_TESTS=YES ..
cmake --build . -j --target cub.cpp14.catch2_test

Expected behavior

Successful build of CUB tests with / without RDC.

Reproduction link

No response

Operating System

No response

nvidia-smi output

No response

NVCC version

No response