Closes https://github.com/cython/cython/issues/4637
See See https://github.com/cython/cython/issues/3556

* Acquire the GIL in nogil functions only when strictly needed on function exit, e.g. for cleaning up temp variables from with-gil blocks or adding tracebacks.

See https://github.com/cython/cython/issues/3554

* Make the GIL-avoidance in 7d99b0f0 actually work in nogil functions and not just nogil sections.

See https://github.com/cython/cython/issues/3558