... and that doesn't necessarily mean that malloc() failed.
We do *not* want to evict a heavy-hitter by an unfrequent element.

Note: even the implementation currently in master *did* return NULL,
so some parts of the code were just wrongly returning ENOMEM.

The implementation is now similar to set-associative caches
that x86 CPU use.  Also the API is changed a bit, leading to
slight simplification of our use patterns.

maintainer may wish to choose neutral or even
more favoured server selection of IPv6 servers.

-DFAVOUR_IPV6=0 makes neutral server selection
-DFAVOUR_IPV6=40 gives IPv6 40ms better chance

RTT tracking for all targets is also supported,
but no loadbalancing is done based on that yet

when doing fast retransmit, each address may be
contacted 1-N times, but previously only cumulative
RTT was tracked for the NS that sent the answer.

now the approximate query start time is subtracted
from cumulative RTT to give an idea how long it
took since the query was actually sent, and all
the NSs that didn't respond within their retransmit
windows are penalised too

example:

> daf.add 'forward 127.0.0.1@5353'

* simplified soft-fail per-ns limit to per-query
  limit, each query gets 4 tries at resolving
* instead of locking at single servfailing NS,
  penalise it and run reelection, this may or
  may not try other servers but avoids pathologic
  case when single NS is servfailing while others
  are good but never probed
* added new nsrep update mode (addition)

this is not going to be backwards compatible change, but it will be the first tagged libknot release sufficient for resolver

any answer that is considered as malformed/servfail/otherwise bad
penalizes the NS for the next time like timeout, this doesn't apply for
DNSSEC validation failures as it still may be okay for insecure
resolution. EDNS failures are okay because the server is requeried in
the most simple RFC1035 mode before flagging it as failed

this avoids instant requeries for SERVFAILing resolutions

if the NSLIST already has a candidate with addresses, never replace
current leader with an empty one. otherwise if the former leader address
was tried and failed, new empty leader would be blamed for this and
penalized

this allows daemon to try other NSs for fast retransmit if the best chosen NS doesn’t respond within time limit

effectively enables/disables usage of given IP protocol
for subrequests (the server can still listen on these)

the selected address is now an array with selection, caller can then send the same query to multiple offered targets

refs #35

libuv doesn't do connected UDP sockets, so we can't get ICMP unreachable
otherwise

previously if NS had no A/AAAA records, they we’re looked up in subrequests, after that a new NS was reelected (possibly preferring another unknown), now it stays true to the NS of choice and changes only if it is unuseable

this LRU-like cache tracks lame nameservers, unresolvable
nameservers (to not waste resources in trying to resolve them),
and possibly other features (extension support, …)

when a NS is in the TIMEOUT, it can’t be autoselected by default but only probed after a successful dice roll - however in some scenarios we’d like to probe timeouting servers more thoroughly 

simplified bee foraging strategy, by default each query goes to the most rewarding NS
however with a chance proportional to the NS RTT, a server may be probed even if it isn’t the most rewarding

score has now meaning of ‘RTT’, maximum RTT is 10s
which is also the penalty for timeout
unknown servers are favorized as 10ms servers to
encourage resolver to try them out, if they contain
unknown glue they are most favourable

nameservers with addresses are preferred, the rest
has same weight

no blacklisting, penalization or rtt scoring yet