Restrict the T attribute to two-dimensional arrays

[kgryte]

This PR:

• restricts the T attribute to two-dimensional arrays in agreement with the mathematical definition of the transpose. For array instances which are not two-dimensional, the specification recommends an error be thrown.
• includes a note explaining the rationale for limiting to two-dimensions and adds a recommendation to use the permute API (note: the permute specification will be added in a subsequent PR; see gh-247).
• removes unnecessary blank lines found in various notes.

Background

Discussion as to why the current behavior of T attribute is problematic can be found in gh-228.

By limiting the attribute to two-dimensional arrays, we retain the current behavior for .T for the primary use case of taking the transpose of a matrix, while removing the less desired behavior of reversing all axes.

[rgommers]

The array instance must be two-dimensional.

(making the same comment twice today, xref pytorch/pytorch#64180): gh-228 seemed to say to raise an error for >2-D only. Cc @lezcano

[kgryte]

Yeah, I am a bit conflicted here. Yes, taking the transpose of a 1D array could be a no-op, but, as discussed in numpy/numpy#7495, it is questionable whether this is actually what you want. In which case, opted to be more restrictive in limiting to only 2D arrays.

[kgryte]

Based on discussions in the consortium meeting held 09-02-2021, no objections were made to limiting to 2-dimensional arrays.

While consensus determined that the transpose of a 1-dimensional array is ambiguous (does a user expect a no-op or actually want to convert a row vector to a column vector or vice versa?), less ambiguous is the behavior of taking the transpose of a 0-dimensional array (a no-op).

However, the stance of this PR is that a user should be explicit in taking transpose of an empty two-dimensional array, rather than using a zero-dimensional shorthand, as not clear to a reader whether an author actually intended for a no-op or this is a bug. Accordingly, strictly limiting to 2D and not allowing 0D avoids this runtime ambiguity.

[kgryte]

Merging based on discussions during the 09-02-2021 consortium meeting.