gh-125498: Replace ghccc with preserve_none in JIT builds

[savannahostrowski]

• Issue: Replace ghccc with preserve_none in JIT builds #125498
• Issue: Update JIT compilation to use LLVM 19 #123714

[savannahostrowski]

CI failures seem to be related to #125521

[brandtbucher]

Awesome! Should I run benchmarks?

Initial review of these changes below. Also, this should probably have its own blurb.

[savannahostrowski]

Once CI passes, yes! Let's run the benchmarks. Thanks for the review!

[mdboom]

Can you merge in main to get this fix in? That way we can benchmark on Windows as well.

[savannahostrowski]

@mdboom Done!

[savannahostrowski]

So it looks like the Python version installed with brew is failing transitively...doing a bit of spelunking here.

[savannahostrowski]

Okay, after a bit of digging here, my cursory diagnosis around the failing x86 macOS runs is essentially this (and I am by no means a Homebrew or Ruby expert):
- The default Python version of Homebrew was recently updated to 3.13 (see Homebrew/homebrew-core#182840). As part of this PR, the formula for Python 3.13 seem to vastly differ from 3.12...and 3.12 was also updated.
It seems like Homebrew is no longer handling overwriting files in the way it once did. More specifically, the combo of the setup-python step and then brew installing Python (via LLVM) is not cooperating. I did some testing here, and it looks like this issue only happens with 3.12 onward (preliminary run of https://github.com/savannahostrowski/test_repo_python_install/actions/runs/11453722971 but waiting on the fail-fast: false run for the full output).

Update as of 10/22:
I got a reply from a Homebrew maintainer: Homebrew/homebrew-core#195288. It looks like this is a known GHA issue that has existed since at least May and started popping up recently as a result of some updates that GitHub made to runner images. The workaround is to remove the symlink to the pre-installed Python so that Homebrew Python is used instead. They may also start investigating overwriting the symlinks when in GHA context on their end because this seems to be a prevalent issue.