CI: Expand test matrix to cover more CUDA minor versions

[leofang]

This might require changing the way we define the matrix (in ci-gh.yaml) now that we need build/test matrices to diverge.

[leofang]

(This task only concerns minor versions; the "major version" parallel is tracked in #275.)