TST: CI instability (xref upstream breakage from setuptools+numpy+astropy)

[neutrinoceros]

Bug report

Bug summary

yt/visualization/tests/test_fits_image.py started failing on certain platforms a few hours ago, see for instance
https://github.com/yt-project/yt/actions/runs/3504488379/jobs/5870201144

It seems that the on-demand-import wrapper around astropy is crashing the test with

ImportError: This functionality requires the astropy package to be installed.

However astropy is indeed installed in those jobs (see the logs). There were no recent changes (less than 24h) in the test itself or the fits_image.py module, astropy wasn't upgraded since the last successful job, and I haven't been able to reproduce this problem locally thus far.

[neutrinoceros]

So, another library that I maintain, unrelated to yt, began failing tests with a suspiciously similar error:

E       ImportError: Convolution C extension is missing. Try re-building astropy.

And pulling the thread leads to pypa/setuptools#3693
This issue affects a lot of projects downstream of numpy, so the best course of action for us is probably to just wait it out.

[neutrinoceros]

That said, if yt on-demand-imports mechanism didn't swallow the initial ImportError message completely, it would have saved me a couple hours of vain research, so I guess that's an easy actionable on our side that could close this issue (when CI is stable again)

[neutrinoceros]

a final note for now: in Python 3.11, we could leverage this feature: https://docs.python.org/3/whatsnew/3.11.html#pep-678-exceptions-can-be-enriched-with-notes

This would be useful if we can manage to keep things tidy on earlier versions too.

[neutrinoceros]

The situation just got resolved with the release of setuptools 65.6.1
I'll keep this ticket open still, and close it with a PR to improve how error messages are displayed when importing extra dependencies fails.