In the following topology

A --[v4]-- B --[v6]-- C

if B does not support v4ov6 (eg. because of an old kernel), it can
nevertheless announce a v4-over-v6 route to C, since the route in its
own routing table is pure v4.

Detection is based on the kernel version, and can be overridden by
the configuration option `v4-over-v6 {true|false}`. If v4-over-v6 is not
supported, ignores v4-over-v6 routes early.

An empty IPv4 source prefix is represented by a v4mapped
default prefix.  This was not always the case.  Thanks to
Fabian Blaese.

The code was removed in 0111f5c1.

The original code was taken from netlink's manpage, which is buggy.
The buffer might not be aligned to struct nlmsghdr pointers.

This would cause an assertion failure whenever we
sent a Unicast Hello.  Thanks to Dave Taht.

  * clarify code flow, only checking for future state once
  * clarify error messages when MTU is guessed by including the MTU value

As suggested by Christof Schulze.

It's a minor optimisation, but it makes me happy.

If no Hello has been buffered yet, insert an unscheduled Hello.
This only works if rfc6126-compatible is false.

This just implements the sending functions, it doesn't actually do anything
with them yet.

This fixes the case where a neighbour only sends unscheduled Hellos.

Incorrect message size when neighbour is not link-local.

Now that we have sub-TLVs and crypto keys, the previous value is no longer
suitable.