Acquire job row locks in a consistent order to prevent deadlocks (#89)

[AudeCstg]

PR to follow changes used by tags

[luca-montaigut]

Fixes #89.

Taking over this PR from @AudeCstg, context in #58 (comment). It is rebased on 26_ux and reduced to what is still needed there: the deadlock retry Aude originally added is covered by the transient-error backoff that _maintain_updates gained in 26_ux (DeadlockDetected is an OperationalError), but the lock ordering is not, and without it a push of existing unique keys overlapping a batch of job updates still deadlocks and fails one side.

What changes

• push_many_ex / sync_push_many_ex insert jobs sorted by unique_key, and still return the references in the caller's order.
• Worker._maintain_updates sorts its pending updates by (unique_key, id) before the executemany.
• Both sort through one helper, chancy.utils.lock_order_key, so the rule cannot drift between the two code paths.
• tests/regressions/test_89.py replays the interleaving deterministically (one transaction holds the first row in worker order while a concurrent push submits the same keys in reverse order). It fails with DeadlockDetected on 26_ux and passes with this patch. A second test checks the reference order of sync_push_many.
• Changelog entry under 0.26.0.

Context

We run Chancy in production at Ouihelp (8 queues, ~60 job types, cron on ~50 schedules) and have been running this fix since April with no deadlock since. Happy to adjust anything.