Add newaxis constant for use in array indexing

[kgryte]

This PR

• adds the constant newaxis to the specification, as an alias for None. This change was discussed and supported in the March 24, 2022 consortium meeting.
• addresses readability concerns when performing array indexing (e.g., A[:, newaxis] vs A[:, None], where the former more explicitly conveys that the selection tuple results in an array with an additional dimension).

Notes

newaxis as an alias for None enjoys broad support among array libraries. Only PyTorch and Dask lack explicit support.

• NumPy: np.newaxis
• PyTorch: no equivalent, but supports None
• MXNet: mx.np.newaxis
• TensorFlow: tf.newaxis
• CuPy: cupy.newaxis
• Dask: no equivalent, but supports np.newaxis (i.e., None)

[asmeurer]

This looks good, and is absolutely trivial for libraries to implement. Obviously this needs to be merged alongside #408. Is the plan to defer this until after the spec release?

[kgryte]

@asmeurer As discussed during meetings, we'll merge this for inclusion in the current specification version given that indexing with None is already universally supported across array libraries.

Most libraries already have a newaxis constant, except Dask and Torch. Torch still needs to add all currently specified constants. Dask has plans to add all constants (dask/dask#8931).

Accordingly, now that #408 is merged, we can merge this PR, as well, for inclusion in the current specification version.