attempt to fix (finally?) the joingingfile logic

the joiningfile is now made by whichever job gets to this task first, so the error where it wasn't fully made (when jobs were slowly run from the HPC queue) should not happen now unless it takes a long time to write that one short file. I guess we could put in a checkfile if required, but this should be *very* rare.

fix some bugs in restarting from snapshots: the start_at was calculated incorrectly (needed to take into account modulo and previous start_at)

all other files access the HPC API, not the slurm and condor objects directly