Further work on #1773 "Loading a session with features extracted from multiplexed inputs". Re-read the aggregate wave models from the session file; also re-order the way they are written so as to improve the likelihood of successfully re-reading them (! - as it stood before, there was some chance involved)