Special case when NaNs present in statistical functions

[honno]

Adds a special case for the two functions that any NaNs present in computed elements should result in NaN. Hopefully I've used the correct language here.

Per discussion with @kgryte, this is expected behaviour.

>>> min([0., float("nan")])
0.0
>>> np.min(np.asarray([0., float("nan")]))
nan
>>> torch.min(torch.as_tensor([0., float("nan")]))
tensor(nan)

[rgommers]

Do note this does mean PyTorch is wrong.

I don't see that in the example. tensor(nan) (as a 0-D array) looks like the expected answer to me (reduction over 1-D array -> 0-D array). The numpy.array_api answer should match that, but it may not (?).

[honno]

Do note this does mean PyTorch is wrong.

I don't see that in the example. tensor(nan) (as a 0-D array) looks like the expected answer to me (reduction over 1-D array -> 0-D array). The numpy.array_api answer should match that, but it may not (?).

Oops I was thinking of a different discussion—I didn't even read my own example! PR description updated accordingly.

[honno]

@kgryte I suppose this special case applies to all statistical functions? If that's the case, I can update this PR.

[kgryte]

@honno Yes, that is correct. NaN values should propagate for the other statistical functions. Feel free to update this PR. Thanks!

[kgryte]

LGTM. Thanks, @honno!