Fix a few race conditions in the state calculation

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

synapse/storage/state.py

@@ -237,6 +237,20 @@ class StateStore(SQLBaseStore):
 
         defer.returnValue({event: event_to_state[event] for event in event_ids})
 
+    @defer.inlineCallbacks
+    def get_state_for_event(self, event_id, types=None):
+        """
+        Get the state dict corresponding to a particular event
+
+        :param str event_id: event whose state should be returned
+        :param list[(str, str)]|None types: List of (type, state_key) tuples
+            which are used to filter the state fetched. May be None, which
+            matches any key
+        :return: a deferred dict from (type, state_key) -> state_event
+        """
+        state_map = yield self.get_state_for_events([event_id], types)
+        defer.returnValue(state_map[event_id])
+
     @cached(num_args=2, lru=True, max_entries=10000)
     def _get_state_group_for_event(self, room_id, event_id):
         return self._simple_select_one_onecol(