Explicit rules about zero-dimensional tensors

[arogozhnikov]

Documentation mentions in several places zero-dimensional tensors, but it's not clear to me if those are required to be supported.

While it sounds very reasonable to assume that e.g. xp.sum(x) returns scalar, some of frameworks don't provide support for zero-dimensional arrays, and fake it with 1-element 1-dimensional arrays. Examples:

• mxnet fails on creation mx.nd.zeros([]) I know API will be reworked in mxnet 2, but did not check actual support there
• paddle also lacks zero-dim arrays, it does not seem to be in any priority

[rgommers]

Thanks for pointing that out @arogozhnikov! 0-D arrays are required, so we should make that more explicit.

[kgryte]

@rgommers Not immediately clear to me where we should put this guidance in the spec. You have any thoughts on this?

We could make a separate "scalars" document and include the requirement there that scalars must be supported as zero-dimensional arrays.

[rgommers]

I'm not sure we need a whole separate section, just a note near the top of https://data-apis.org/array-api/latest/API_specification/array_object.html should be fine I think?

[kgryte]

Okay. Will add.