Release v19.91.2

Fixed

• A catch-up over many partitions no longer fails steps whose events were already delivered, or leaves the catch-up job running forever, when the per-step handled-event reports queue up on the observer. Reports are now cheap in-memory updates written together, a report resent after a timeout is not counted twice, a late report no longer fails a handled step, and any other step failure records a failed partition so it is retried. (part of #4954)