Ensure allow_threads without requiring the GIL

[wyfo]

I encountered recently the issue of needing allow_threads in Drop implementation. It forces me to write Python::with_gil(|gil| gil.allow_threads(|| {...})).
In the context of Python code executing the drop, this is fine, because the GIL will already be acquired, but if my object is somehow used in Rust code, dropping it will require acquiring the GIL while it's completely useless (and can cause performance issues)

Maybe PyO3 could expose some allow_threads free function, which will not try to acquire the GIL, just check that it's not acquired or release it if needed.

I didn't open a issue because I don't need this feature myself, as my objects should not be used in Rust code, but this feature may interest you, especially as it is quite easy to implement (and you know I'm always volunteer to do it myself if needed.).

[adamreichold]

That seems like a good rationale and I think if we do commit to #3646 which would change the API for allow_threads in any case, we should go for a static method Python::allow_threads or free function allow_threads immediately.

SuspendGIL has to access GIL_COUNT in any case and checking the stashed value against zero and becoming a no-op in that case should have no measurable overhead at all.