Due to chunked lazy sending of megolm sessions it was in theory that we encrypted two olm
messages to the same device in different futures out-of-order. Introducing locking here should
fix this (increadibly rare, so far only theoretical?) race-condition