Fix crash in gil_scoped_acquire

[rostan-t]

Description

Increase gilstate_counter before calling PyThreadState_Clear in gil_scoped_acquire to prevent finalizers to try to create and destroy a new thread state while the current thread state is being cleared.

Fixes #5827

Suggested changelog entry:

Fix crash that can occur when finalizers acquire and release the GIL.

[rwgk]

I just added this a couple days ago (PR #5822), to git rid of very frequent failures under macOS / free-threaded that nobody had the time to debug:

pybind11/tests/test_gil_scoped.py

Lines 205 to 209 in 852a4b5

	if env.MACOS:
	if not env.sys_is_gil_enabled():
	pytest.xfail(f"[TEST-GIL-SCOPED] {soabi} with GIL disabled: " + msg)
	if env.PY_GIL_DISABLED:
	pytest.xfail(f"[TEST-GIL-SCOPED] {soabi}: " + msg)

WDYT about removing that in this PR? — So that we see immediately if your fix helps with that, too. If not, we can put that code back in a follow-on PR.

[rwgk]

Could you please move this somewhere around here?

pybind11/include/pybind11/detail/common.h

Line 97 in 852a4b5

I think PYBIND11_HAS_STD_BARRIER would be ideal (similar to the PYBIND11_HAS_STD_LAUNDER that we have already).

[rwgk]

Totally not sure: Is there any point in guarding this with #if defined(...) for certain Python versions and Py_GIL_DISABLED?

[rostan-t]

I focused on free-threading here but I think that the issue can also occur in destructors of thread-local variables in the regular build.

[rwgk]

In other places we use a pattern that would be defined_PYBIND11_HAS_STD_BARRIER here.

(It makes it a little easier to pin-point all things connected to PYBIND11_HAS_STD_BARRIER across Python and C++ code.)

[colesbury]

FWIW, this LGTM