_auth_and_persist_outliers could do with being optimised

[richvdh]

As part of my work on #9595, I've ended up making some of the code which persists events received over federation a bit more conservative. (Specifically, this is the code for handling batches of outliers, in FederationEventHandler._auth_and_persist_outliers).

#10896 (comment) raised the concern that this could result in a lot of round-trips between the federation inbound worker and the event persister.

It should be possible to optimise all this, so that we do all the event auth first, and then persist everything in one big chunk.

[richvdh]

The main gotcha here is that we do not currently set event.rejected_reason on events until they have been persisted and reloaded. Instead, we currently set a flag in EventContext, which check_auth_rules_for_event doesn't consider.

It might be worth considering getting rid of EventContext.rejected, and just setting event.rejected_reason. The two are a source of confusion. (For example: I think that the condition at

synapse/synapse/handlers/federation_event.py

Line 1009 in f563676

if marker_event.rejected_reason is not None:

will always be false.)