Large batch of to-device messages was dropped

Logs show that a batch of (150) to-device messages received over federation was stored by the `encryption` worker, but never sent out to users by the synchrotrons. Obviously this led to unable-to-decrypt errors.

It looks like the batch was correctly INSERTed into Postgres, but the replication logic failed to find it, and instead of sending out an `RDATA` with the necessary data, we instead sent an empty `POSITION` command. The synchrotrons therefore did not attempt to read the messages from the database, and ended up just deleting them when the user made the next sync request.