more attempts to clean up the code to work better on NFS : seems ok now, but expect more "bug" fixes to come (and more cleanup as the code is a bit hacky)

few other cleanups

add automatic cacheing of various functions, and a cache_test() routine to test the runtimes with and without caching

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