GH-140815: Acquire GIL or stop the world when calling _Py_DumpTracebackThreads from faulthandler_thread

[sergey-miryanov]

Mostly a PoC now.

skip_current_tstate was added to prevent the modification of tests that use the check_dump_traceback_later function. The faulthandler_thread is now in the traceback and has no Python frames, so the output of those tests may change. This is okay, as the faulthandler_thread doesn't have one.

• Issue: SEGV in faulthandler.dump_traceback_later #140815

[vstinner]

I don't think that this change is correct. By design, faulthandler can be called on a deadlock. It should be able to dump the Python traceback without acquiring any kind of lock since Python may already been blocked somehow.

I wrote a different fix #140921 to detect invalid/freed frame or code object.

[sergey-miryanov]

Yeah, if we should avoid any locking at all costs then this fix is not correct. I will close this one and take a look at your #140921.

Thanks!