This is not a complete fix for #125174, just enough to unblock other work

🤖 New build scheduled with the buildbot fleet by @markshannon for commit 94f8fd0 🤖

If you want to schedule another build, you need to add the 🔨 test-with-refleak-buildbots label again.

The two "failing" refleak buildbots seem to be hanging in test.test_concurrent_futures.test_shutdown. There are no reported refleaks.

Performance impact:

Overall a small slowdown, with a large 7-10% slowdown on nbody.
This is almost certainly due to the regression in stack spilling around _Py_IsImmortal_Loose fixed here.

ARM 64 darwin
x86-64 linux

I'll re-run the linux benchmarks with the fix as a sanity check.

Windows shows an over 2% improvement. With the regression fixed, this I would expect this to be more than 3%.

This reason for this is that the MSVC does not convert unsigned int x; x > (1 << 31) into signed x; x < 0 as do clang and gcc.
This PR uses the signed x; x < 0 form. Hence the speedup.

x86-64 linux performance numbers are a bit noisy, but show no significant change.