Reduce the cost of freeing replica systems at checkpoints #253

Profiling large replica exchange systems shows that a large part of each checkpoint is spent freeing the copy of each replica's system once its files have been written, and on clearing its trajectory frames afterwards. Both hold the GIL, so replicas can't checkpoint in parallel during this time. This issue tracks freeing these systems in a way that lets the checkpoint workers overlap, using new support in Sire, and clearing the frames in an order that avoids an extra copy of every molecule.