Fix off-by-one in code which checks for over-long domain names
in received DNS packets. This enables buffer overflow attacks
which can certainly crash dnsmasq and may allow for arbitrary
code execution. The problem was introduced in commit b8f16556,
release 2.73rc6, so has not escaped into any stable release.
Note that the off-by-one was in the label length determination,
so the buffer can be overflowed by as many bytes as there are
labels in the name - ie, many.

Thanks to Ron Bowes, who used lcmatuf's afl-fuzz tool to find
the problem.

Only in DNSSEC mode, where we might need to validate or store
such names. In none-DNSSEC mode, simply don't cache these, as before.

Return INSECURE when validating DNS replies which have RRSIGs, but
when a needed DS record in the trust chain is proved not to exist.
It's allowed for a zone to set up DNSKEY and RRSIG records first, then
add a DS later, completing the chain of trust.

Also, since we don't have the infrastructure to track that these
non-validated replies have RRSIGS, don't cache them, so we don't
provide answers with missing RRSIGS from the cache.

  If multiple routes to the same network exist, Linux blindly picks
  the first interface (route) based on destination address, which might not be
  the one we're actually offering leases. Rather than relying on this,
  always set the interface for outgoing unicast DHCP packets.

The nasty code with static variable in retry_send() which
avoids looping forever needs to be called on success of the syscall,
to reset the static variable.