Fix issue #211

[lohedges]

Closes #211. _save_energy_components() now records each row in a per-window buffer and the new _flush_energy_components() writes the buffered rows in one append from the base _checkpoint(), which both runners go through and which already holds the checkpoint lock. Previously every energy save called parquet_append(), which reads the whole file and rewrites it, costing 3.8 ms per replica per cycle on a 46646 atom system against 90 ms of dynamics at 1 ps cycles, and growing with the length of the run. The components file is now consistent with the other checkpoint files, where before it was written outside the lock and could run ahead of the checkpointed energies; a crash loses at most one checkpoint interval of components, as it already did for the energies. Restart deduplication is unchanged, since it reads the file's last time on the first save after a restart. To bound memory when checkpoints are rare or disabled, the buffer is also written out whenever it reaches 10000 rows, which is a few megabytes and one extra append per 10000 saves.

Three tests are added in tests/runner/test_energy_components.py: one runs both runners for 12 fs, restarts to 24 fs, and checks the file holds one row per energy save with no duplicates; one counts calls on the RepexRunner and checks six saves produce two flushes and six rows; and one lowers the row limit and checks it triggers the extra flushes without duplicating rows.