When picking destination servers that we need to wake up for a retry, we need
to be mindful of destinations that we have *never* successfully sent to. This
can manifest either as a null `last_successful_stream_ordering`, or even no row
in `destinations` at all.

Hence, we need to left-join on `destinations` rather than inner-joining, and we
need to treat a null `last_successful_stream_ordering` the same as 0.

When considering which destinations need waking up for a retry, also look for
those that have outstanding to-device messages.

We don't really want people to have to wait 60 seconds for their to-device
messages.

…vice_messages