we should kill off the event_reference_hashes table

[richvdh]

as far as I can tell, the values we read from it are never used. They are dutifully passed all the way down the stack into create_new_client_event, which throws them away.

[richvdh]

so I started on this in #6629. It turns out that event_reference_hashes is still used when building V1 events (ie, events for room_versions 1 and 2), so we can't kill it off just yet.

[richvdh]

Thinking about this a little more: I still think we should kill off event_reference_hashes: we can replace the hash lookups for v1 event population with calculations. The reasons for this are:

• most events are (or soon will be) not v1 events, so calculating and storing the reference hash is wasteful
• of those that are v1 events, most are only used once or twice in references, so again: storing their hashes is wasteful. (The exception is things like the create event and PL event of busy rooms, but caching the calculation result will mitigate this)
• Calculating reference hashes isn't that hard and if it saves waiting for a database lookup it's probably a win anyway
• Event persistence (as opposed to creation) is a bottleneck since it's hard to parallelise, so it's good to move any work we possible can out of the event persistence path.