fix optimistic-lock conflicts in update/delete and make fjall cas atomic

[fabracht]

Summary

• update/delete now retry on optimistic-lock conflict, converging to field-level last-writer-wins; the typed Error::Conflict remains the backstop after a bounded number of attempts.
• Fix a non-atomic CAS in the fjall backend (check-then-act → silent lost update under concurrency). The commit's precondition check and apply are now serialized; fsync stays outside the lock.
• Version bumps: mqdb-core 0.7.3, mqdb-agent 0.8.8, mqdb-cli 0.8.9.

Test plan

• cargo make dev (format + clippy-pedantic + full suite) green
• new concurrency tests: update convergence and delete/index consistency
• before/after bench (--durability none, concurrency=1): no write-throughput regression