To-do following from #13754: we may be able to remove cuda-version entirely and make packages that only depend on (arch, Python version) and not CUDA version.

We would only build packages on one CUDA version entry in the matrix (per Python version, per arch), and test them across all Python/arch/CUDA combinations. I think custreamz and dask-cudf are the only candidates for reduced builds like this. I think cudf_kafka may not be a good candidate because it links to CUDA C++ libraries, but I'm not sure.