Make sure that CUDA(-dependent) builds are always done in an accelerator prefix #823

I accidentally triggered a CUDA+UCX-CUDA build for zen2 without accelerator target in #817 (comment), and this worked fine, so I almost overlooked that I overlooked that I forgot to specify the accelerator target. It would be nice to have some check (CI, hook?) in place that prevents us from doing that, i.e. to make sure that these CUDA builds always are done in an accelerator prefix.