fix(python): Don't allow passing missing data to generalized ufuncs

[itamarst]

Fixes #14811

This is a more constrained PR, just addressing one specific issue: passing missing data to generalized ufuncs can silently lead to incorrect results.

Generalized ufuncs are a NumPy feature (https://numpy.org/neps/nep-0020-gufunc-signature-enhancement.html), but the easiest way to implement them is using Numba, which is what the tests do.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 80.91%. Comparing base (9bfa30c) to head (1f176ce).
Report is 9 commits behind head on main.

Additional details and impacted files

@@            Coverage Diff             @@
##             main   #16198      +/-   ##
==========================================
- Coverage   80.99%   80.91%   -0.08%     
==========================================
  Files        1392     1394       +2     
  Lines      178930   179574     +644     
  Branches     2904     2913       +9     
==========================================
+ Hits       144925   145310     +385     
- Misses      33502    33759     +257     
- Partials      503      505       +2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[MarcoGorelli]

Implementation looks good to me, just got some quite minor comments

[MarcoGorelli]

Thanks @itamarst !

[itamarst]

Thank you @MarcoGorelli. If you have commit access you'll have to merge it too, since I can't.

[MarcoGorelli]

yup - just leaving it open another day in case anyone has comments, then I'll ship it

outside of generalized ufuncs, there's no other case where missing values are a concern, right?

[itamarst]

For normal value-by-value ufuncs, if you have memory that is e.g. an invalid float64 then that could potentially be a problem? You're potentially feeding random garbage to the function for missing data.