Fix AS sender catch up handling, after an AS comes back online by batching up events in transactions

This issue has been migrated from [#8704](https://github.com/matrix-org/synapse/issues/8704).

---

If the AS is marked down then for every event persisted we create a new transaction for that AS with a single event in it. When the AS comes back online Synapse will then send these transactions one at a time, which can mean that an AS can take a very very long time to catch up, even if its processing the events quickly. We should send transactions with multiple events instead (though not too many).

(Split out from #8604)

The issue names no files, tests, or entry points. Start by locating Synapse's application-service sender catch-up handling and determine how persisted events are turned into transactions; done means an AS returning online receives bounded transactions containing multiple events rather than one event per transaction.

Written by the indexing model from the issue text.