server selection: consider switching to TCP instead of backing off the timeouts to high values

The following discussion from !1030 (merged) should be addressed:

• @sbalazik started a discussion: (+1 comment)

  config.hints test is timing out sometimes on this branch and so far, I have no idea why.

  22/36 knot-resolver:postinstall+config+skip_asan / config.hints  TIMEOUT 120.05 s 
  --- command ---
  KRESD_NO_LISTEN='1' PATH='/builds/knot/knot-resolver/.local/sbin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin' TEST_FILE='/builds/knot/knot-resolver/modules/hints/tests/hints.test.lua' SOURCE_PATH='/builds/knot/knot-resolver/tests/config' /builds/knot/knot-resolver/tests/config/../../scripts/test-config.sh -c /builds/knot/knot-resolver/build_ci/../tests/config/test.cfg -n
  --- stdout ---
  /builds/knot/knot-resolver/.local/sbin/kresd
  processing test file /builds/knot/knot-resolver/modules/hints/tests/hints.test.lua
  ok 1 - has IP address for a.root-servers.net.
  ok 2 - load root hints from file
  ok 3 - can retrieve root hints
  ok 4 - real IP address for a.root-servers.net. is replaced
  ok 5 - real IP address for a.root-servers.net. is correct
  [65536.00][rplan]   [qry tree] badname.lan. A (0) <- 
  [65536.00][rplan]   [push] pending 1; badname.lan. A (0) | resolved 0
  [65536.03][rplan]     [qry tree] . DNSKEY (3) <- badname.lan. A (2) <- 
  [65536.03][rplan]     [push] pending 2; . DNSKEY (3); badname.lan. A (2) | resolved 0

This is because the iter_ns_badip.rpl workaround allows the pushing of the same query to rplan twice in the row which leads to multiple tries with back-off of the timeout to resolve . DNSKEY or a.root-servers.net AAAA (if DNSSEC is turned off). The old selection implementation switches to TCP after a few tries and there the connection fails and the NS address is flagged as 'bad'.

Switching to TCP instead of backing off into big timeouts might be a good idea which might even help with the pathological cases that appear in respdiff now.