asarray partial fix

[ivirshup]

Attempt at a fix for asarray.

• array_api_compat.numpy.asarray(torch.Tensor) returns a torch.Tensor #106

Opening as a draft in case someone else gets the opportunity to look at this earlier, as some cases still error. I think the test should be useful to anyone else attempting this.

[asmeurer]

I'm not sure what the right fix here is. The code was correct before the __array__ check was added in 8ee1613, but I'm not completely clear what the purpose of that check was.

Also, this code seems to deal mostly with copy logic, and that logic needs to be refactored, similar to data-apis/array-api-strict#15 (at the very least, copy=False should work in numpy 2.0). This version of the function is going to be more complicated than the one in array-api-strict though because it also has to support things outside of the standard (i.e., any valid input to np.asarray), and also because it's a wrapper function that's reused for numpy, cupy, and dask.

[asmeurer]

+1 to this test. We are going to need to add a lot of tests here for asarray, partly because the test suite only handles things within the scope of the standard, and partly because the test suite itself doesn't test as much as it could (data-apis/array-api-tests#241). We can also adopt the tests from data-apis/array-api-strict#15 (once we have fixed the copy behavior, but I expect that will need to just be part of this overall fix since this is part of that logic).

[asmeurer]

This was included with #110