Improve performance of remove_{hidden,deleted}_devices_from_device_inbox

[babolivier]

See #11401 for context

We do this by using the batch size to cap the stream IDs rather than using it as a limit.

[richvdh]

le sigh. I tried this on postgres - it's trying to table-scan devices for me, which is prohibitively expensive.

try this for size:

DELETE FROM device_inbox di
WHERE
    stream_id >= ? AND stream_id < ?
    AND NOT EXISTS (
        SELECT * FROM devices d WHERE d.device_id=di.device_id and d.user_id=di.user_id AND NOT hidden
    )

You'd think that NOT EXISTS would be the same as (key) NOT IN, but clearly not. Sorry, there's no great magic here - I just tried various ways of writing the query until I found one where postgres decided to use the index on devices.

Note that it does both updates in one pass. The idea is we want to filter out non-hidden devices from the inbox.

[richvdh]

nearly there!

I think we could probably have gotten away with just turning remove_deleted_devices_from_device_inbox into the new update, but this is fine too.

[richvdh]

lgtm otherwise