Federation retries

turn back on per-request transaction retries, so that every time we try to hit a dead server we actually end up hammering 5 times :|

add errbacks to enqueue_pdu deferreds; change logging for failed federation sends to warn rather than exception

track replication destination health, and perform exponential back-off when sending transactions. does *not* yet retry transactions, but drops them on the floor if waiting for a server to recover.

Store any incoming presence push in the local cache anyway, even if there's no interested observers (yet *hint*) (SYN-115)

Workaround for non-uniqueness of room member events in the database confusing HAVING COUNT() test of room membership intersection (with thanks to Tom Molesworth)

Squash room_ids list to a set() to remove duplicates - with TODO marker as I have no idea /why/ the dups are happening in the first place