This removes duplication between unicast and multicast variants of message
formatting.

This is no longer useful and complicates the code.

Thanks to Christof Schulze for noticing the problem.

When adding unreachable routes and setting the RTNH_F_ONLINK flag, a
device is required to be specified. In Linux kernel 4.16 support for
this flag was added. Until now it was ignored.
If RTNH_F_ONLINK is specified while the device is missing, newer kernels
will respond with No such device.
The result is:
* spam in the log file
* missing routes for both ipv4 and ipv6

The check for a TLV going beyond the end of the packet was off by two.
A malformed packet could possibly cause babeld to read two octets beyond
the end of the read buffer.

While technically a buffer overflow, this is most probably not
exploitable, since it is a read-only overflow.  At worst, it would
cause two octets of garbage to be parsed and treated as valid data.

This fixes a bug that was introduced in commit 4f4e3cba, and prevented
non-source-specific IPv4 routes from being redistributed.  Thanks
to Niklas Yann Wettengel for the detective work.

Point-to-point interfaces are bound to two link-local addresses: the
local address and the peer address [1].  The former is advertised with
an IFA_LOCAL TLV and the latter with an IFA_ADDRESS TLV.

[1] $ip addr show [...]
    inet6 fe80::1234 peer fe80::5678/128 scope link

Since send_request was buggy, these weren't doing anything.  Don't change
the behaviour, sending wildcard requests at startup is not a good idea.

When we lose a route and have no feasible alternate, we should send requests
straight away rather than waiting for a periodic update.  If we have an
unfeasible route, send according to that route and resend, otherwise do
a multicast.

Reported by janus.