Since this is a rather costly operation, we only do it when we overflow
the source table.

Instead of neigh->id[0] = 0xFF.

This is more taken care of in update_neighbour_metric.

This avoids the send_self_retract hack.

This ensures that retractions are sent too.

This avoids receiving stale data in answer to the update.

Now that requests work well, there's no need to increase seqno_interval
as fast, and this makes it more likely that routes will be feasible.