Be a bit more careful about how we calculate the state to be returned by
/sync. In a few places, it was possible for /sync to return slightly later
state than that represented by the next_batch token and the timeline. In
particular, the following cases were susceptible:
* On a full state sync, for an active room
* During a per-room incremental sync with a timeline gap
* When the user has just joined a room. (Refactor check_joined_room to make it
less magical)
Also, use store.get_state_for_events() (and thus the existing stategroups) to
calculate the state corresponding to a particular sync position, rather than
state_handler.compute_event_context(), which recalculates from first principles
(and tends to miss some state).
Merged from PR https://github.com/matrix-org/synapse/pull/372
fddedd51d9