Add `release_gil_before_calling_cpp_dtor` annotation for `class_` (#5522)

* Backport of https://github.com/google/pybind11clif/pull/30088 (main PR) and https://github.com/google/pybind11clif/pull/30092 (minor fixes).

Note for completeness:

These are identical to the current versions on the pybind11clif main branch (@ commit 4841661df5daf26ecdedaace54e64d0782e63f64):

* test_class_release_gil_before_calling_cpp_dtor.cpp
* test_class_release_gil_before_calling_cpp_dtor.py

* Fix potential data race in test_class_release_gil_before_calling_cpp_dtor.cpp

The original intent was to let the singleton leak, but making that tread-safe is slightly more involved than this solution. It's totally fine in this case if the RegistryType destructor runs on process teardown.

See also: https://github.com/pybind/pybind11/pull/5522#issuecomment-2661068351

---------

Co-authored-by: Ralf W. Grosse-Kunstleve <rwgk@google.com>