Setup Cirrus CI

[nbraud]

• pytest on the same platforms as ppb-vector
• Added a docs/requirements.txt file
• Build docs in a separate task.
• Install pygame's build dependencies
• Only install build dependencies if no binary wheel is available.

[nbraud]

@pathunstrom I forgot to leave a comment yesterday, but this should be good-to-go; the Windows failure is the timing issue that already exists on master (though I'd totally understand waiting for that to be fixed so you can check Cirrus goes green)

[nbraud]

@astronouth7303 I just noticed there already was a requirements-docs.txt, so I just removed docs/requirements.txt.

Do you think it would be reasonable to rename that file though? I didn't expect requirements-*.txt (and instead *-requirements.txt or */requirements.txt), hence why I didn't notice it, and I have the impression the latter is more common than the former (test-requirements, dev-requirements, ...)

[nbraud]

I'm experimenting with caching dependencies in custom container images in nbraud#1

[nbraud]

Update: I got the test time (for Linux tasks) down to ~2s in nbraud#1, from resp. ~10s (if there's a pygame wheel) and 2 minutes (when there is no wheel).

I'm going to try and see if I can easily make the same thing for Windows containers.
Would you prefer that work in this PR directly, or in a follow-up PR? (I'd prefer the latter, but it's no strong preference)

[pathunstrom]

We can definitely do that as a follow on pr.

And @astronouth7303 there's that windows failure again.

[pathunstrom]

@nbraud You comfortable with this? It seems to be functioning and I'd like to get it in since it'll make our lives better on the rest of the work getting done.

[nbraud]

PS: I forgot to mention yesterday, but I believe nbraud#1 is ready. I did generalise the Dockerfiles to Windows containers, and got the test time there down to ~5s. I didn't yet address the issues we discussed on Discord about automatically picking up new versions of dependencies from PyPI, and I think that would involve using entirely different tooling, so that's basically it for “v1” of dependency caching.

macOS doesn't support anything like containers, so the same approach won't work there (though @astronouth7303 mentioned that it might be possible to persist the pip cache and perhaps even a whole virtualenv)

[pathunstrom]

Is nbraud#1 supposed to be against your fork?

[nbraud]

Yeah, I didn't want to add all of it to this PR right away, but I still wanted to have Cirrus run it; I'll open a new PR from that branch once this lands, I guess.