Here is my first draft of the DNS TCP fallback support that seems to
be basically working. There may of course be bugs, and there are cases
where I probably don't want it to fall back like it's doing, but it's
at a stage where review and [smoke-]testing would be helpful.

Some known issues:

- getaddrinfo should use larger buffers (probably 640-768 bytes) or
  the "huge CNAME" problem iss not solved.

- If res_send is used to send EDNS0 queries crafted by the caller and
  the caller provides a buffer smaller than what the OPT header
  advertises, the answer will be silently truncated to the caller's
  buffer size.

I've tested querying TXT for locations.publicdns.goog (>7k RRset) and
basic non-fallback usage. Further large-answer query test cases would
be appreciated, especially names where one or both of the A or AAAA
RRset are large so that fallback in parallel lookup can be tested, and
later so that getaddrinfo buffer size increase can be tested.

Rich