Support testing np and tnp on the same test

[honno]

@ev-br here's how one could test both NumPy-proper and torch_np on the same test. Essentially any test case just needs a parameter np, which should be treated as the array module we want to test (i.e. numpy, torch_np), and we need to ensure we're using that module rather than outside imports.

I reworked some of test_indexing.py to demonstrate how this could be utilised. It's a bit finnicky as you have to check np is {numpy,torch_np} for some module-specific behaviour, e.g. raised error classes and skips/xfails, and you have to make sure you're using the argument np for things like the {numpy,torch_np}.testing utils.

Maybe we could support np as a magically param but not really bother to rework existing tests with it—instead we use it as-and-when we feel it useful to "sanity check" a test works the same for NumPy. So IMO it might be worth merging just the conftest.py changes.

[ev-br]

Merging conftest.py change (+explanation, this looks rather black magic to me at least :-)) LGTM. The rest, let's wait for Mario to weight in?

[ev-br]

Thank you @honno!
Let's track desired patterns of actually using this feature in #78