Reduces initial sync times from ~30s to ~1s on accounts with heavily
populated rooms.
The problem was that f.e. RoomMember it would try to calculate the
display name, which involved looping each RoomMember to get their
display name to check for disambiguation. We now cache display names
to user IDs so we don't need to loop every member when disambiguating.
The RoomState needs to be current before higher-level processing as this
processing may depend on the end state rather than the progressive state
(e.g. disambiguating BOTH RoomMember's display names).
New sentinels are only created when the RoomMember state changes, so we don't
needlessly deep copy RoomMembers f.e. MatrixEvent. Sentinels co-exist with
RoomState.members which are single instances to which listeners can be attached.
This gets the best of both worlds (don't have to keep re-attaching listeners on
member changes, don't have needless memory consumption).
Accessed via MatrixEvent.sender property. Deep copy the list of state events
from initial sync for old/current RoomState so updating the .sender property
doesn't affect both of them. Reverse the insertion of initial sync MatrixEvents
so state diverges to *earlier* points in time. Add a 'forwardLooking' property
to MatrixEvent to determine which out of 'content' and 'prev_content' entities
should take into account e.g. when determining display names. Finally, always
create new RoomMembers when they are updated in order to prevent corrupting
(read: sharing the same member object) the MatrixEvent.sender property of
existing events in the timeline.