TST: Add dev numpy to devdeps, fix compat with numpy 1.24

[pllim]

This might smoke out the error that #151 is trying to fix.

I also see this error in https://github.com/astropy/astropy-integration-testing/actions/runs/3604333353/jobs/6073585087

Possible suspects:

• BUG: Make mask_invalid consistent with mask_where if copy is set to False numpy/numpy#22046
• DOC: Add changelog for masked_invalid change. numpy/numpy#22406

[codecov]

Codecov Report

Merging #153 (ca1dc16) into main (6b2ffb0) will decrease coverage by 0.45%.
The diff coverage is 50.00%.

❗ Current head ca1dc16 differs from pull request most recent head 58434f7. Consider uploading reports for the commit 58434f7 to get more accurate results

@@            Coverage Diff             @@
##             main     #153      +/-   ##
==========================================
- Coverage   75.77%   75.32%   -0.46%     
==========================================
  Files           9        9              
  Lines         706      697       -9     
==========================================
- Hits          535      525      -10     
- Misses        171      172       +1     

Impacted Files	Coverage Δ
specreduce/core.py	73.17% <50.00%> (-1.19%)	⬇️
specreduce/extract.py	94.70% <50.00%> (-0.36%)	⬇️

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[pllim]

Hmm. Maybe #151 isn't the correct solution after all.

[ojustino]

Hi @pllim, after some local testing, I think the best solution would be to drop masked_where() entirely. In the place that causes the error you found, all we really need is the mask; the other attributes that come with a masked_array are superfluous. All they do to create the mask is call isfinite(), so we can just do the same here.

This usage of masked_where() appears a few places in specreduce and I believe they should all should be changed in the same effort. That could be this one, or if you like, I can try it in a different pull request.

[pllim]

If you want to change mask logic, it is up to you as I am not familiar with the algorithms here. I just did track down why numpy dev is giving error and it might or might not be upstream bug. I am going to ask numpy people (see numpy/numpy#22720).

[pllim]

As for the coverage, it will fix itself when numpy is released and your coverage job picks it up. Obviously it is covered because tests fail with dev numpy.

[pllim]

Maybe we can merge by the end of the week if numpy makes it clear that they have no intention of backing out this behavior.

[ojustino]

@pllim I opened #155 to make the edits I talked about earlier since most of the places that trigger this issue don't actually need to use mask_invalid(). Do you have any objections to that approach?

[ojustino]

The other PR should probably still integrate the changes you made to tox.ini in order to catch these errors more quickly in the future.

[pllim]

@ojustino , yes, please do cherry-pick the dev numpy job over. If #155 supersedes this PR, feel free to close this without merge. Thanks!