When deaggregating on redacting an annotation relation, we remove the
relation and aggregate the other relations for that key again, so we can
reliably detect the first timestamp (and count and me as well to lesser
extent).
as a consequence, more than one event can get updated when redacting a
relation (the relation is updated, as well as the relation target), so
account for that by returning an array of entries that have updated.
We'll need to to aggregate whether we have reacted to a message
Create writers at room level and pass subwriter is dependency, rather
than creating them in sync and gap writer.
