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.

Fixes https://gitlab.labs.nic.cz/knot/resolver/issues/93.
API of a KR_EXPORT function is changed, so ABIVER is bumped.

This saves a few kilobytes in the executable.
Also, the name XXD seemed no longer suitable,
as it does lua-specific cleaning.

and set libknot_SONAME and libzscanner_SONAME as lua literals.

Remove now obsolete libpath lua function - use find_soname from
platform.mk to define <arg>_SONAME, add lua_pushliteral to
daemon/engine.c and add -D to daemon/daemon.mk for any new library
loaded from Lua.

The cookie module could segfault due to the `opt` pointer not being
initialized when no EDNS was sent in the incoming request.

Based on [issue 28](https://github.com/CZ-NIC/knot-resolver/issues/28), we were able to boost kresd performance running in a Docker container on DockerCloud by nearly 100% by binding to a single interface:ipaddress. It might be handy for other users to know.

As noted in #94, it feels natural to call it like:
``net.listen({net.lo, '192.168.1.1'})``
Also, minor fixes were done in that function and corresponding docs.

if the score is known (and not glued), the retry
rate is <avg(rtt), 250ms>

if the NS score is unknown or worse than 250ms,
it is always retried at this rate

all other servers in the list use default rate

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

Consequently, there's no use for worker_reserve in the header.

Also break overlong lines.

The last hunk: we'd report worker failure even if we didn't
attempt to start it due to previous failures.

We called kr_crypto_init() just before forking and then _reinit() in
all children, but we never did any cryptography in-between - better
initialize after forking.

Note: the function is KR_EXPORT.  It might not be useful anymore, but
its removal would require bumping libkres soname.

timeouts over TCP when <n> first nameservers timeout over UDP as
previously the TCP would connect again to the first and only the
first nameserver in task->addrlist.

This would need to be ultimately fixed in TCP Fast Retransmit.

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'

The missing entry caused some entries to be dear and written to wrong
positions.