`device_lists_changes_in_room` is inefficient.

[erikjohnston]

We see a lot of disk IO on the matrix.org DB go on reading from the device_lists_changes_in_room table.

I think this is due to the fact that we have a background process that processes each row, and then flips a boolean flag on the row (which is in an index). The causes the row to effectively get deleted/reinserted, meaning that a lot of the recent (ish) rows get shuffled around causing many more pages to have to be loaded from disk to service queries.

My proposal is to do one of:

1. Have a separate table to keep track of devices that we need to process, and then delete out of that when we've processed a row. The disadvantage with this approach is that we double the amount of data we need to insert; or
2. Remove the boolean flag and instead keep track of which stream ID we had processed up to. This has the disadvantage that we need to scan over every row of device_lists_changes_in_room table even though we may not need to handle them (i.e. we only need to process local device changes).