The previous fix for the ongoing event fetches counter
(8eec25a1d9) was both insufficient and
incorrect.
When the database is unreachable, `_do_fetch` never gets run and so
`_event_fetch_ongoing` is never decremented.
The previous fix also moved the `_event_fetch_ongoing` decrement outside
of the `_event_fetch_lock` which allowed race conditions to corrupt the
counter.
Empirically, this helped my server considerably when handling gaps in Matrix HQ. The problem was that we would repeatedly call have_seen_events for the same set of (50K or so) auth_events, each of which would take many minutes to complete, even though it's only an index scan.
