REV: "ENH: Improved performance of PyArray_FromAny for sequences of array-like

[charris]

See #13958.

This reverts commit 71fc59d
which was part of gh-13399 and the followup commit
ba53a63 from gh-13663.

The issue is that we mix Sequence protocol and __array__ usage
when coercing arrays.

[h-vetinari]

This uses the text from the OP of #14109 verbatim (AFAICT), but that is not representative of this PR (eg. milestones, branches), which is quite confusing...

[seberg]

I will have a look at it, the diff here suggest we may already have backported it manually? Since Matti rightliy pointed out there is a big gain here... In my branch, I started caching the __array__ (and list(...) results. But I did not finish this off, and it may look a bit different on current numpy (depends a b it on how Nathaniels feedback turns out).
That caching comes with its own issues, since users could create inconsistencies (although it is possible to add a sanity check at least). But I think it would simplify this issue because it removes the need of having 2.5 code paths doing this.

[charris]

Close/reopen.

[charris]

Putting this in. Improvements should be made on top of this.

[seberg]

@charris was this actually full revision? It seems like something had gone wrong there and it only reverted the test but no other code?

[charris]

Oops. Fixed up in #15027.