fix(voyager): order item popping by timestamp, not id

since we requeue retryable errors with the same id (since
52e1a80fdd52d98f10b554b03aecf0248b160475), if there are many failing
messages (due to a dead rpc, for example), all of the workers will be
used to spin these failing messages. instead of ordering by id when
popping an item from the queue, we now instead order by the created_at
timestamp, and reinsert failed messages instead of simply rolling back
the database transaction.