Synchrotrons always wake up sync streams for presence updates

This issue has been migrated from [#8955](https://github.com/matrix-org/synapse/issues/8955).

---

When calculating presence changes we decide whether we should actively wake up sync streams, or if they should just be bundled up the next time we send a response for another reason. We do not appear to propagate whether to wake up sync streams across replication, and we appear to *always* notify for presence updates:

https://github.com/matrix-org/synapse/blob/6d02eb22dfde9551c515acaf73503e2500e00eaf/synapse/app/generic_worker.py#L381-L390

It's unclear how much this will help.

(Note that we batch up presence updates that we don't notify for and persist/replicate them once a minute).

Start at synapse/app/generic_worker.py lines 381-390 and trace how the decision to wake sync streams is handled across replication. Compare notifying presence updates with updates that are batched and persisted or replicated once a minute. Done means the wake decision is propagated so non-notifying updates do not always actively wake sync streams.

Written by the indexing model from the issue text.