[lohedges]

Closes #250.

Checkpointing was slower than it needed to be. Sire's CPU thread limit applied during checkpoints as well as dynamics, and replica exchange wrote its replicas' checkpoints one at a time.

• Sire uses every CPU thread while checkpoint files are written, unless max_sire_threads is set. The dynamics limit is restored afterwards.
• Replica exchange checkpoints replicas in parallel. Trajectories are still reassembled one replica at a time at the end of the run, which keeps the extra memory to one system's worth.
• The shared GCMC topologies are written once per checkpoint, instead of by every replica.
• The regular runner's worker processes now apply the Sire thread limit, which previously only took effect in the parent process.
• Runners created one after another in the same process no longer inherit each other's thread limits, and a limit set before the first runner is kept.