split_multi_job makes assumptions about the job JSON that are not enforced by the schema

In particular:

 1) an element of the device_groups list is assumed to have tags
 2) the job is assumed to have a logging_level key

(there may be other bits too).

The consequence of this is that submitting a job that fails either of these assumptions just 500s, rather than getting an informative message. Either (a) the scheme needs to be tighter or (b) split_multi_job needs to be more forgiving (and (b) is the correct choice here :-))