We send hellos with a very short lifetime when leaving the network.
This was broken when I implemented automatically adjusting the hello
interval.

Lors d'un appel netlink_read, on suppose que NLM_F_MULTI est toujours
vrai, sauf dans le cas d'un appel depuis kernel_callback.

I don't understand why, but it fixes the problems we have, and Gregoire
says it's okay.

This fixes a bug where we would stick to a suboptimal route if the
best route was very stable.