Avoid data races in BorrowFlag

[ngoldbaum]

Fixes #4904 (specifically the race I found here, the others have been reported upstream).

I haven't been able to see any observable impact of the data races that TSAN finds. That said, with these changes I no longer see race reports from ThreadSanitizer. Note that you need to use Python 3.14 to be able to use thread sanitizer usefully without creating a ton of suppressions, free-threaded CPython 3.13 has a number of races with fixes in 3.14 that probably won't be backported.

As soon as we can add 3.14 in CI I'll look at setting up free-threaded TSAN CI using the cpython_sanity docker image @nascheme has been working on.

I added some comments to explain the choices I made. I'm no expert in C++ atomics though so I'd appreciate some more careful eyes on this code.

I tested this by running this command in the 3.14t docker container after cloning pyo3, checking out the PR branch from #4811 and installing rustup to install a nightly toolchain and the rust-src component:

TSAN_OPTIONS="suppressions=$PWD/suppressions.txt:halt_on_error=1" LD_LIBRARY_PATH="$(pyenv prefix)/lib" LDFLAGS="-L$(pyenv prefix)/lib" RUSTFLAGS="-Zsanitizer=thread" cargo test --lib --test test_gc -Zbuild-std --target x86_64-unknown-linux-gnuTSAN_OPTIONS="suppressions=$PWD/suppressions.txt:halt_on_error=1" LD_LIBRARY_PATH="$(pyenv prefix)/lib" LDFLAGS="-L$(pyenv prefix)/lib" RUSTFLAGS="-Zsanitizer=thread" cargo test --lib --test test_gc -Zbuild-std --target x86_64-unknown-linux-gnu

I created a suppressions file for a race in CPython that's been reported upstream:

root@6221150df46c:/work/pyo3# cat suppressions.txt 
race:long_from_non_binary_base

Ping @davidhewitt.

[davidhewitt]

Sure thing, the stronger guarantees here don't look like they can hurt.

[alex]

Being more conservative can't hurt, and being TSAN clean feels valuable, so this LGTM.

[ngoldbaum]

Thanks for the quick validation! Added to the merge queue...