[lohedges]

Builds on #251, and needs the Sire changes in OpenBioSim/sire#510.

• Free each replica's committed system without holding the GIL, before deleting its frames, so frame deletion doesn't copy every molecule. The regular runner frees it first too.
• Join trajectory chunks into the final trajectory with Sire's new DCD.concatenate, without reloading the topology, so replicas no longer take turns at the final checkpoint.
• Record frame times in final trajectories, rather than zero for every frame.
• Use plain arrays rather than numpy matrices for swap counts.

For a 137,000 atom ABFE system, normal checkpoints went from about 33 s to 12 s, and the final checkpoint from about 300 s to 13 s.