Release v19.13.3

Fixed

• A replay of a projection no longer leaves its MongoDB read model frozen when the replay ends on more than one kernel silo at once, or overlaps another replay of the same projection. The replay is promoted exactly once, and every silo goes back to writing the read model itself. Before, a silo that lost the race kept sending every live change to replay-<container>, with the observer still reporting Active and no error recorded (#4296)
• A replay no longer fails to end because the revert container name from a previous replay of the same read model is reused (#4296)
• Starting a replay on several silos at once no longer fails with a duplicate key on replay-contexts (#4296)