cc @mdboom

Kicking off a benchmarking run now...

The raw benchmaring results are here.

This PR makes things 5% faster on 64-bit Windows, 7% faster on 32-bit Windows, and no measurable change on x86_64 Linux.

The important comparison, however, is whether this change counteracts all of the negative effects of #118450. Comparing directly to commit d611c4c (the commit immediately before #118450 landed), shows the same amount of speedup 5% faster on 64-bit Windows, 7% faster on 32-bit Windows, and no measurable change on x86_64 Linux.

So this PR is effective and seems to make up for the regression.

However... if we could hold off on merging it until we confirm one other thing. @zooba reminded me that we still have a pragma "hack" to disable optimization on MSVC for the main eval loop. This was introduced to work around a compiler crash when the main eval loop function got extremely large (when the Tier 2 and Tier 2 interpreters were merged together). Of course, this hack is no longer needed for non-JIT builds, since they no longer include the Tier 2 interpreter. I am doing another benchmarking run just removing the hack to see if it also fixes the performance issue -- if we can do that and use inline functions rather than macros, that seems preferable.

EDIT: Added link to change that removed Tier 2 interpreter from default builds.

Thanks Mike for the attempt! Sadly, removing the pragma doesn't seem to speed up anything:
https://github.com/faster-cpython/benchmarking-public/tree/main/results/bm-20240702-3.14.0a0-a03affd

As the reported benchmark results are good, and this PR seems to make up for all the perf loss, I will merge this PR, to restore performance of Windows builds.