Add support for expanding dimensions using None

[kgryte]

This PR

• resolves Allowing None for indexing to add an axis #360 by adding support for expanding dimensions via None. The functional equivalent of the proposed changes is repeated invocation of expand_dims. Using None in a selection tuple is more convenient when needing to insert multiple size 1 dimensions.
• as per comment, using None in selection tuples enjoys broad support across array libraries.

[rgommers]

I think adding None makes sense. My main question is whether it's okay to combine it with other indexing semantics. E.g,:

>>> import torch
>>> t = torch.arange(5)
>>> t[None, :3, None]
tensor([[[0],
         [1],
         [2]]])
>>> t2 = torch.arange(6).reshape((3, 2))
>>> t2[:2, None, ...]
tensor([[[0, 1]],

        [[2, 3]]])

That probably needs testing and specifying.

[asmeurer]

I think adding None makes sense. My main question is whether it's okay to combine it with other indexing semantics. E.g,:

There shouldn't be any issues with this I would imagine. The text from the NumPy docs seems fine, although it could perhaps be more precise:

Each newaxis object in the selection tuple serves to expand the dimensions of the resulting selection by one unit-length dimension. The added dimension is the position of the newaxis object in the selection tuple. newaxis is an alias for None, and None can be used in place of this with the same result.

This should probably not apply to boolean indices where we should keep the "only index" rule.

[kgryte]

I've added an explicit note that the use of None in combination with boolean array indexing is not allowed. This should address @asmeurer's comment above.

[kgryte]

Re: combining with other indexing semantics. As @asmeurer mentions, this should not be an issue, given that the use of None is syntactic sugar for first extracting a selection and then applying expand_dims. Meaning, an implementation could filter a selection tuple for non-None values, compute the selection, and then apply expand_dims to the result. As such, there should be no conflict with using None in conjunction with other single-axis indexing expressions.

[asmeurer]

This looks good. Were we going to add newaxis?

[kgryte]

As this PR has been open for 3 weeks, has received approvals here, and no objections were raised during consortium meetings, will go ahead and merge. Any changes can be addressed in follow-up pull requests.