Previously when forked tasks didn't finish synchronously, they were
added to a list of children tasks and destroyed at the next sync.

But this meant that in the case of an infinite loop spawning new forks,
such memory could grow unboundedly and never be reclaimed.

Fixing this also required changing how exceptions in forks are propagated.

- Centralize the logic in scheduler.hpp
- Remove worker.hpp and introduce pool.hpp
- Allow a different number of pools than there are threads