2022.12 support

[asmeurer]

This is still WIP because we need some fixes to the test suite.

[rgommers]

@asmeurer are you also adding fft extension support, in this PR or a separate one?

[asmeurer]

The test suite doesn't check anything yet, but assuming torch.fft and so on already match the spec then there's nothing to do.

[asmeurer]

Thanks to work from @honno at data-apis/array-api-tests#184, I have some updates on pytorch complex numbers support:

The following torch functions do not support complex numbers:

• expm1
• round
• sign
• unique_*

round and sign should be easy to work around. expm1 and unique_* are harder to work around directly so I probably won't attempt that.

[rgommers]

unique

unique_* are harder to work around directly so I probably won't attempt that.

That's probably caused by torch.unique sorting by default and having a keyword for it; sorting won't work with complex dtypes. Not too relevant for the standard, but a pain to get to work indeed in a wrapper. So I agree with your decisions here.

[honno]

LGTM!

Regarding array_api_compat.torch.take(), IIRC index_put() wasn't quite adequate for "full" NumPy compat and I ended up using torch.Tensor.__get__() (Quansight-Labs/numpy_pytorch_interop#96), but it's not that surprising if it's fine for the Array API scope. I'm sure attempted-adoption in SciPy will be a brilliant way to battle-test array_api_compat.torch.take() in the end anywho, just note my test_take does have a few limitations...

https://github.com/data-apis/array-api-tests/blob/f82c7bc8627cc2c3a44fa3e425f53a253a609aa8/array_api_tests/test_indexing_functions.py#L21-L24

[asmeurer]

We figured out @tupui's isdtype issue. It was related to a version of the function that he had implemented in scipy. It seems there aren't any issues with the version in the compat library. So we can go ahead and do a release here.