[CUDA] Only allow NVIDIA offload-arch during CUDA compilation.

Otherwise, if someone specifies a valid AMD arch, we may end up triggering an
assertion on unexpected arch later on.

Differential Revision: https://reviews.llvm.org/D105295