Add a result_type function to the API specification

[rgommers]

This PR adds a specification for result_type, which is the function needed to be able to do explicit dtype casting for portable results in case of mixed integer - floating point operations. Identified as needed in gh-14, also came up in gh-43 and gh-91.

Notes

• Most libraries have a result_type function with the signature proposed here; MXNet is the exception.
• The array-api-comparison tooling says TensorFlow doesn't have it, however in TF 2.4.0 the experimental.numpy namespace does have it.

[rgommers]

@kgryte can you have a quick look at this? I think it's straightforward, it's just the first function spec I add, so want to make sure I've ticked all the right boxes.

[kgryte]

@rgommers One thing to note is that the Torch API for result_type has the signature

result_type( tensor1, tensor2 );

where tensor1 and tensor2 may be either a tensor or number. This is obviously a bit different than NumPy's API.

[rgommers]

Yes I missed that, thanks. The PyTorch version can be extended to the one in this PR without backwards compatibility issues though. The PyTorch one is a bit too limited, it's not uncommon to for example do:

result_type(x, y, complex64)

[rgommers]

Filed a PyTorch issue for it: pytorch/pytorch#51284

[kgryte]

LGTM.