In ReplicatedFastCountManager::initializeMetadata(), we must check if there is a pre-existing watermark during recovery. If yes, the flusher should not emit another watermark until the flusher persists its first checkpoint.

This change is motivated by the pathological case where a node experienced several faults consecutively, emits several watermarks without corresponding checkpoints, and the tailer gets stuck behind. It also simplifies the oplog application logic on standby.

See design for more info.