Python uses wrong randomizer version

[fenhl]

I believe the cause of intermittent failures to recognize some item names in spoiler logs is that there's something wrong with the logic used for importing Python modules from a specific randomizer version. Maybe a race condition between multiple different versions being used at the same time or something. The fix would probably be to cleanly separate Python interpreter instances between randomizer versions. Ideally, this would be achieved using PyO3/pyo3#576, but until that feature is ready, running Python interpreters as subprocesses should be a (less performant) workaround.