Serve repeated requests from the newest finished job, never a failed one (fkie-cad/mcritweb#47)

[r0ny123]

Closes fkie-cad/mcritweb#47 (labelled mcrit there). The cache selection was wrong: a force rematch that was still queued or running shadowed the finished job whose result a repeated plain request could actually use.

What changed

get_cached_job_id picked the newest job with the same descriptor that had attempts left and was not terminated, whatever its state. Both queues now prefer a finished job over an unfinished one and the newest within each group; jobs without attempts left (failed) or terminated ones are never reused, exactly the rule the issue proposes.

• MongoQueue: one query, sorted finished_at descending then created_at descending. A descending sort places every date before null, so the finished jobs come first without a second query.
• LocalQueue: the same rule over its jobs; the _descriptor_to_job map, which could only ever remember the last job, is gone.
• test_job_cache_prio encoded the old rule (a queued forced job wins over the finished one) and is updated; new tests cover finished-over-running, newest-finished, and that failed/terminated jobs are skipped, on both queues.

Verified against a running instance

MongoDB 7 + this branch as server and worker, four samples, a finished 1vs1 job between two of them:

step	result
force rematch, then cancel it, then plain request	served by the earlier finished job
force rematch still running, plain request meanwhile	served by the earlier finished job
forced job finished, plain request	served by the newer finished job

ruff, ty and the full suite (mongo tests included) pass.