Define a "job" for the ML and light-HPC systems

[lars-t-hansen]

See here for a general discussion of this matter. We need to define what a "job" is for systems without a job queue, because the "job" is the most reasonable unit of work that we'll be dealing with in the other tools. Most reasonably a job is something that is synthesized from the output of sonar. Sonar is effectively a sampling profiler, providing a snapshot of running processes with their current resource usage. A simple tool that monitors that output may be able to build up a set of "jobs" over time and may log them in the same form that SLURM logs them. It's possible that we need to augment the output of sonar a little bit so as to distinguish multiple back-to-back runs by the same user of the same programs as separate jobs - this should not be too hard.

[lars-t-hansen]

A sensible system has been implemented for this in sonar, and seems to work OK.