Use Postgres state and explicit sync waits in the crash-recovery test

[jmqd]

This PR makes the randomized crash and restart test more realistic and reduces potential flakiness.

• Stores replication progress in Postgres and reopens it after the simulated crash, rather than
  reusing in-memory state.
• Waits for initial table synchronization before testing crashes, and for the table to be ready
  after restart before shutting down.
• Clarifies that recording an event at the destination does not mean the pipeline has finished
  processing its response.

The test still checks that replication recovers without copying the table again. No production
behavior changes.

[coveralls]

coverage: 76.549% (+0.004%) from 76.545% — test/postgres-store-dirty-restart into main