gh-140795: Put the thread state in PyCriticalSection structs

[encukou]

GH-141406 improved performance by only fetching thread state once and storing it in a variable on the stack.

This instead puts the thread state in the PyCriticalState struct (also a temp variable on the stack), bringing the public and private implementations closer together.

• Issue: performance regression in ssl module under free-threading #140795

[colesbury]

I used Claude to benchmark the PR on Linux in a few cases (main executable, Py_BUILD_CORE extension, non-stdlib extension). tl;dr it seems a 2-3% slower in microbenchmarks of Py_BEGIN/END_CRITICAL_SECTION and essentially no difference in larger benchmarks.

https://github.com/colesbury/gh-140795-benchmarks/blob/main/results.txt

[kumaraditya303]

improved performance by only fetching thread state once and storing it in a variable on the stack.

The important part that this PR misses is that in my PR the thread state is passed in a register whereas here it will need to be loaded from stack, I don't see value in doing this unless the performance is atleast equal.

[encukou]

I see. In my (much more limited) testing this looked better.
Sorry for the noise!