fix: bump to manylinux 2010, add test

[henryiii]

From #157. I've uploaded manually built 1.11.1 binaries for manylinux2010. This adds a test that shows the breakage, then adds the bump to 2010 which fixes it. Should we also ship manylinux1? I'd say no (in fact, I think I'd be fine to delete the old manylinux1 binaries for 1.11.1) - it's probably more likely that and old pip would get a broken binary than someone would need this on actual manylinux1.

I'd like to know why this broke - was is ninja, or is there some issue with the latest manylinux1? Maybe someone will investigate someday.

FYI, to build locally with a different manylinux, say for the 32-bit wheel:

git checkout 1.11.1
CIBW_MANYLINUX_I686_IMAGE=manylinux2010 pipx run cibuildwheel --only cp39-manylinux_i686

[mayeut]

It's probably more likely that an old pip would get a broken binary than someone would need this on actual manylinux1

yep, for the past week that's around 5200 linux downloads with an old pip (<19.0) versus 50 downloads with 2.5 <= glibc < 2.12
Overall, that's 0.8% of downloads.

We could probably drop manylinux1 now. If the lack of a manylinux1 wheel is a problem, then a mix-in with manylinux1 wheels built with a statically linked libstdc++ (as done in #159 ) can be added.

I'd like to know why this broke - was is ninja, or is there some issue with the latest manylinux1? Maybe someone will investigate someday.

Given that manylinux1 has been EOL for some years now, nothing has changed (for the relevant issue). My best guess is that a code change in ninja triggered an already existing bug in devtoolset-2 (rather than libstdc++ breaking ABI compatibility which might still be a possibility).

[mayeut]

@henryiii, since you approved #159, I merged that one and will close this one.