Add CPython 3.10 wheels

[hrnciar]

PR Summary

This PR modifies cibuildwheel.yml to create wheels for Python 3.10. I propose this as a draft because it is currently blocked on numpy's wheel (numpy/numpy#19630).

PR Checklist

• [N/A] Has pytest style unit tests (and pytest passes).
• [N/A] Is Flake 8 compliant (run flake8 on changed files to check).
• [N/A] New features are documented, with examples if plot related.
• [N/A] Documentation is sphinx and numpydoc compliant (the docs should build without error).
• [N/A] Conforms to Matplotlib style conventions (install flake8-docstrings and run flake8 --docstring-convention=all).
• [N/A] New features have an entry in doc/users/next_whats_new/ (follow instructions in README.rst there).
• [N/A] API changes documented in doc/api/next_api_changes/ (follow instructions in README.rst there).

[github-actions]

Thank you for opening your first PR into Matplotlib!

If you have not heard from us in a while, please feel free to ping @matplotlib/developers or anyone who has commented on the PR. Most of our reviewers are volunteers and sometimes things fall through the cracks.

You can also join us on gitter for real-time discussion.

For details on testing, writing docs, and our review process, please see the developer guide

We strive to be a welcoming and open project. Please follow our Code of Conduct.

[QuLogic]

I guess there's no point to use manylinux1 for Python 3.10 now?

[hrnciar]

Yes, Python 3.10 does not support manylinux1 and it was failing with it.

[henryiii]

There's actually no point to use manylinux1 for 3.9 either - the last reason for manylinux1 was the weird hack of attaching pip 9 to python3.8-dev on Ubuntu 18.04. NumPy does not ship manylinux1 wheels for 3.9 anymore.

[hroncok]

The 3.10 wheels for numpy are manylinux2014. I suppose here we should do the same?

[tacaswell]

Sorry for the draft thrashing, I was optomistic that the numpy issue being closed meant they had uploaded wheels already.

[hrnciar]

@tacaswell numpy 1.21.2 was released and wheels are available. My PR works for Linux and Windows wheels, but I am getting an error on macOS build. It seems to me like an error in Numpy, WDYT? https://github.com/hrnciar/matplotlib/runs/3338097720

[henryiii]

numpy 1.21.2 was released and wheels are available. My PR works for Linux and Windows wheels, but I am getting an error on macOS build.

Only 64-bit linux (Intel and ARM) wheels are included. There are no Windows or macOS wheels. NumPy on macOS requires special setup because it can't use the buggy and deprecated built-in Accelerate framework.

[rgommers]

NumPy on macOS requires special setup because it can't use the buggy and deprecated built-in Accelerate framework.

Not quite true anymore, Accelerate was re-enabled in numpy.distutils recently after some work by Apple. But we still use OpenBLAS for wheels. And it will not be re-enabled for SciPy.

[QuLogic]

I think this will need a rebase.

But while this is likely to pass a build, it does not mean that tests are passing, due to problems in our dependencies, see #20864 (comment)

[tacaswell]

Put on the 3.5.0 milestone so we get auto backporting when this is merged.

[QuLogic]

If we skip macOS, we can get wheels on Linux and Windows.

However, as noted above, NumPy only builds manylinux2014 wheels for 3.10. This means that our build somewhat needlessly rebuilds NumPy, and we should probably switch that to build manylinux2014 only.

[QuLogic]

I skipped macOS and bumped 3.10 wheels to manylinux2014: https://github.com/QuLogic/matplotlib/runs/3489831873 There don't appear to be 32-bit wheels for NumPy, so those still get rebuilt, so not sure if that saved anything.

[hroncok]

Let's skip 32bit wheels as well? They can always be added later.

[QuLogic]

I mean, they build, just a bit slower; no need to skip them.

[tacaswell]

@QuLogic if you think this is ready to go please self-merge.