Implement "everything follows X" for GridSearchCV - #33633

Refreshed this PR. It is more of a small fix now then really implementing something new.

Could be an easy review.

cc @lucyleeow @virchan

@betatim there is one failing test. Could you please check?

The failure happened because we use y.dtype.metadata['unique'] as a place to store unique values for numpy arrays. This short cut is cool, but it turns out when you convert a numpy array (with such metadata) to a cupy array the metadata is kept, but the type of the cached metadata isn't changed. So later on you end up mixing cupy and numpy arrays. I think the reason we run into this is because the two calls to attach_unique were not removed. It looks like github.com//pull/34221 remove all the other attach_unique calls but missed these two. So I removed them. This seems to fix things.

I added a new combination of "numpy" -> "cupy" to the mixed namespace test generator. This would have caught this bug before. But I could be convinced that we are starting down a slippery slope of combinatorics by doing this. The yield_mixed_namespace_input_permutations function is explicit about the fact that it does not want to yield all possible combinations because that would be too much. I think I'd add this case, because cupy does not strip the metadata, but torch does.

I added a new combination of "numpy" -> "cupy" to the mixed namespace test generator.

I think this is very reasonable. When we made the initial list, it was a best guess at what would encompass "all" cases.
Only thing I would think about is maybe adding a note in the yield_mixed_namespace_input_permutations docstring, about it also including a case of 2 different namespaces but both have the metadata attribute (for attach unique).