The basic idea is that we load the Python bytecode once and then fork to spawn all the needed worker processes, thus reducing CPU contention.

Empirical attempts show that this reduces CI time of Complement-with-workers by a half.

It's an attempt at making #12848 better.