Simplify detection of re-entrant initialization of pyclass types

[davidhewitt]

          The detection of re-entrant initialization seem surprisingly heavy weight. What do you think about turning this around in the future, meaning to use e.g. a thread-local list of type object addresses which the current thread is currently initializing instead of a list of threads for each type object?

Originally posted by @adamreichold in #2947 (comment)

[adamreichold]

This does not work as reentrancy can happen both from the current thread which could be detected using a thread-local list and from other threads at the same time, so we still need global synchronization. I guess, this could be done using hazard pointers but that seems a bit too much to reduce start-up time.