Run CuPy tests on CI

[asmeurer]

We need to run the CuPy tests on CI somehow. Right now, I have to manually run CuPy tests locally when I do a release. But this only works because I have a local machine that has a CUDA GPU. If someone else wanted to do a release, they would need to access such a machine. But also, this means that we only really check CuPy when we do a release. I also check it manually when doing a CuPy related PR, but having it run more often would be nice. It's also easy to forget to do this when releasing.

There are ways of getting GPU runners for CI, which we should use here.

[betatim]

How about using GitHub hosted runners with a GPU? Or are there other "CI with GPUs" that you know of/this project could use?

[rgommers]

The other option is using Cirun and using an AWS/GCP instance. Also not hard to set up, although GHA is slightly easier (or just more familiar at least). The main thing is the "connect a credit card and manage spend" issue, which is the same either way. I'd go with GHA here I think, and the label-based triggering that you set up for scikit-learn @betatim.