Remove _CCCL_HAS_CUDA_COMPILER()

[davebayer]

I've removed almost all uses of _CCCL_HAS_CUDA_COMPILER() and now it's the time to remove the macro completely. Basically it's just almost always not the right tool for the job.

Most of the uses were/are replaced by _CCCL_CUDA_COMPILATION() and the rest was/is almost always replacement for _CCCL_CUDA_COMPILER(NVCC) or _CCCL_HAS_CTK().

The compiler detection is now wrapped by _CCCL_CUDA_COMPILATION(), so if we are not compiling CUDA code, there is not CUDA compiler. This was not always the case, because if the user does:

$ nvcc -x c++ my_src.cpp

we would still set the _CCCL_CUDA_COMPILER_NVCC() even though we are compiling a c++ file. In my opinion this is not correct, because nvcc just controls the compilation trajectory and doesn't compile any source code.

This means that many of the builtins that cudafe++ doesn't support are not used even though the c++ compiler could use them without any problem.

So, this PR removes _CCCL_HAS_CUDA_COMPILER() completely and puts all compiler/compilation macros to a consistent state.

[fbusato]

I strongly agree with the motivation of the PR. _CCCL_HAS_CUDA_COMPILER() is way too generic. I want to double-check we are not introducing subtle errors.

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 14h 21m: Pass: 100%/91 | Total: 2d 10h | Max: 2h 58m | Hits: 83%/206947

See results here.