Combine test workflows into one

[michaelosthege]

This PR combines pytest, arviz-compat and jaxtests workflows into one (big) tests workflow.

Motivations

• Cancelling all workflows is now 3x less clicking. (Only core devs have the necesary permissions to cancel workflows.)
• GitHub Actions workflows can't depend on each other, but if we want to fail fast on ❌ pre-commit (Adjust pre-commit step in CI to run every time, but fast fail if red #5041), having all jobs in one workflows will make it a lot easier.
• Tracking workflow runtime via https://share.streamlit.io/koaning/gitlit/main will be easier after they're all in one.

Changes

• ✔ Removed the buildosx build script. Chris confirmed via Slack that it can be removed.
• ✔ arviz-compat and jaxtests were deleted - their jobs were moved into tests (previously pytest.
• ✔ Upgraded the tests from ubuntu 18.04 to 20.04 which is the most recent LTS release.
• ✔ Redistributed tests between jobs to reduce the runtime (closes Combine windows workflows #5475)

[codecov]

Codecov Report

Merging #5623 (c3c383f) into main (849a64d) will decrease coverage by 0.03%.
The diff coverage is n/a.

❗ Current head c3c383f differs from pull request most recent head a9955ac. Consider uploading reports for the commit a9955ac to get more accurate results

@@            Coverage Diff             @@
##             main    #5623      +/-   ##
==========================================
- Coverage   87.67%   87.63%   -0.04%     
==========================================
  Files          76       76              
  Lines       13717    13717              
==========================================
- Hits        12026    12021       -5     
- Misses       1691     1696       +5     

Impacted Files	Coverage Δ
pymc/smc/smc.py	96.38% <0.00%> (-1.61%)	⬇️
pymc/util.py	76.25% <0.00%> (-1.25%)	⬇️
pymc/sampling.py	85.97% <0.00%> (-0.12%)	⬇️
pymc/distributions/simulator.py	87.50% <0.00%> (+1.38%)	⬆️

[ricardoV94]

Looking at the runtime there are 4 or 5 workflows that took ~6 minutes to run, one of which only spent 45 seconds running actual tests, and the rest was setting up the environment

https://github.com/pymc-devs/pymc/runs/5612971573?check_suite_focus=true

I would merge those short workflows with some of the others that take like 11 minutes. We still have some workflows that take ~30 minutes, so we will not loose anything by merging the fast ones.

[michaelosthege]

I aimed for 15-20 minutes per job. Only the test_distribution.py still takes much longer. But I don't want to split it up in this PR.

I was about to comment final success, but then I noticed that the Windows jobs were only running the first item in each subset 😮 And that was already the case on main 😲
I home my latest push fixes it, but in any case I'll probably have to do more balancing.

Some stats compared to 849a64d:

• Number of test jobs reduced from 23 to 22
• Total runtime reduced from 6h38m in to 6h22m

[ricardoV94]

LGTM,

This also explains why the windows xfails were not failing with strict XPASS in the past

[ricardoV94]

Failing test seems to be #5575

I'll try to fix it this week. Shouldn't be a blocker here

[michaelosthege]

I noticed that the test_distribution_moments.py takes 12 minutes, which is much longer than I would have expected.

My last force-push moved mixture tests between the ubuntu groups to shave the longest job runtime down.

I'm going to rebase-merge since there's no need to run the CI again.