[wip] limited api wheels

[minrk]

try building wheels with limited API

requires Cython 3.1, currently in prerelease

see #2010 for more about which wheels we could/should build

• I think we should definitely drop CPython 3.8-3.10, leaving those to the Stable ABI 3.7 wheel
• The most we could drop is everything but 3.7 and 3.12 (and 3.13t)

closes #2010

[minrk]

Actually can't drop any wheels before 3.11 because 3.11 introduces Py_buffer to the ABI, which we need. So we still need to build wheels for each supported Python <=3.10.

So we can't drop most wheels, but we could drop CPython >= 3.13 or 3.12 in favor of a limited API wheel. It's unclear if the vectorcall improvement in 3.12 is worth it or not.

[minrk]

cibuildwheel doesn't build wheels if there's already a compatible one, so if we want to build both a 3.11 abi3 wheel and a 3.13 native wheel, we need to do something special with ci (i.e. build abi3 wheels in their own jobs).

Running with Python 3.13, I saw a ~20% performance penalty of about 1µs per send (baseline 4.5µs) in the throughput benchmarks with the 3.11 abi (basically just measuring the overhead of Socket.send). the penalty was reduced to about 0.7µs for the 3.12 abi with vectorcall.

So I can either:

1. publish a stable abi wheel for 3.11, stop building newer wheels (~1µs penalty)
2. publish standard wheel for 3.11, stable for 3.12+ (~0.7µs penalty)
3. keep publishing all 'current' wheels, but publish just the 3.12 stable abi, so that future Pythons can have a wheel before PyPI allows uploads (more wheels, not fewer)

I'm not sure which I'll pick when the time comes, but perhaps option 2.

I really hope freethreading gets a stable ABI in 3.14, too.