TEST: Only use stable argsorts in PARREC tests

[effigies]

numpy/numpy#23707 introduced an optimized argsort that will be used on some platforms, including GitHub actions' runners, making this a nasty issue to figure out. I'm very glad that it wasn't on some obscure architecture that one user hit, as I have no idea how I would have debugged that.

---

Between the last good and first bad runs, numpy released 1.25 and GitHub Actions released new Python images. First starting by verifying that the previous versions of both resolve the issue. Then will relax numpy.

[codecov]

Codecov Report

Patch and project coverage have no change.

Comparison is base (b7022e0) 92.16% compared to head (c385a53) 92.16%.

Additional details and impacted files

@@           Coverage Diff           @@
##           master    #1234   +/-   ##
=======================================
  Coverage   92.16%   92.16%           
=======================================
  Files          98       98           
  Lines       12364    12364           
  Branches     2539     2539           
=======================================
  Hits        11395    11395           
  Misses        646      646           
  Partials      323      323           

☔ View full report in Codecov by Sentry.
📢 Do you have feedback about the report comment? Let us know in this issue.

[effigies]

Okay. So it is Numpy 1.25. Still unclear why it's 100% of the time here and I've been unable to reproduce even stochastically on my local machine.

[effigies]

Looking through the commit logs between the last succeeding and first failing pre-releases: numpy/numpy@126b46c...v1.25.0

I think numpy/numpy#23707 is a likely candidate. Will try finding alternatives to argsort for the failing tests.

[effigies]

@grlee77 Would you mind having a look? Does this seem like a thing worth reporting upstream, or is instability in argsort considered to be acceptable? (Nevermind. I'll go ahead and report and they can close it if they like.)

[effigies]

Bug report here: numpy/numpy#24064

We'll see if we need to include this patch or if we should just tolerate failures until numpy releases a fix.

[effigies]

Nevermind. This was our fault for not explicitly requesting a stable argsort. Merging.