Release v19.26.2

Fixed

• A projection replay in a cluster no longer loses the last writes of the silo it ran on: every silo writes what it holds back before any silo ends the replay and promotes the rebuilt read model (part of #4448)
• A silo that finds a replay already ended by another silo stops writing to the replay collection, so later changes reach the read model instead of a collection nothing reads (part of #4448)
• Two replays of the same read model within one second no longer fail promoting the second one with "target namespace exists" (part of #4448)