Like #132956 this only helps -S startup because io gets imported anyway by site.py and possibly other places. Maybe we can make those other places also use _io but not convinced it's worth it.

I'm going to run this branch on our benchmarking hardware (as well as #132956), just so we have some numbers.

Benchmarking of this PR: https://github.com/faster-cpython/benchmarking-public/blob/main/results/bm-20250425-3.14.0a7%2B-dac50cc/bm-20250425-linux-x86_64-JelleZijlstra-sunder_io-3.14.0a7%2B-dac50cc-vs-base.svg

This is actually more effective than (19%) than #132956 (12%). I think it makes sense to merge this, and maybe the other just for good measure (though it should have no additional effect).

This is now ready.

One tricky aspect is that we need to check any side effects that come from importing io. Most of the file is defining ABCs and registering things to those ABCs; it's fine if we omit that because the ABCs don't exist if the module isn't imported. However, it also sets the __module__ attribute on the UnsupportedOperation exception. I changed the code to set that attribute in _io instead because otherwise users who never import io would see this exception with the wrong module name.

When you're done making the requested changes, leave the comment: I have made the requested changes; please review again.

I have made the requested changes; please review again

Thanks for making the requested changes!

@mdboom, @gpshead: please review the changes made to this pull request.