forks are connected with IPC pipes to process
group leader and can execute expressions on its
behalf. so running commands over all workers
is easy now:

> hostname() -- single
localhost

> map 'hostname()' -- all
localhost
localhost
localhost

this change introduces new API for cache backends,
that is a subset of knot_db_api_t from libknot
with several cache-specific operations

major changes are:
* merged 'cachectl' module into 'cache' as it is
  99% default-on and it simplifies things
* not transaction oriented, transactions may be
  reused and cached for higher performance
* scatter/gather API, this is important for
  latency and performance of non-local backends
  like Redis
* faster and reliable cache clearing
* cache-specific operations (prefix scan, ...) in
  the API not hacked in
* simpler code for both backends and caller

* daemon now processes messages over TCP stream
out-of-order and concurrently
* support for TCP_DEFER_ACCEPT
* support for TCP Fast-Open
* there are now deadlines for TCP for idle/slow
streams (to prevent slowloris; pruning)
* there is now per-request limit on timeouts
(each request is allowed 4 timeouts before bailing)
* faster request closing, unified retry/timeout timers
* rare race condition in timer closing fixed

an internal timer walks RTT timer periodically and
clears entries with bad results every 5 minutes.
this means that a timeouted entry penalty is 
capped to that interval, making sure that the
bad reputation doesn't last forever

resolver will always attempt to contact upstreams
known to be bad if it's not busy.
this fixes a problem on low-volume resolvers
where a short connection outage could make
resolvers deny resolving queries even after the
connection is restored

breathe failed to process the typedef thinking
the macro expansion was a function pointer

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

this doesn’t guarantee valid cache after crash,
but then it’s a cache. most of the time on cold
cache is spent on fsyncing as it’s done per each
commit (=> resolved query)

an average request from cache requires ~12k, an average iterative request 16-64k
this reduces allocation cost when there are now pools on the freelist

when the daemon starts in non-interactive mode, it creates a pipe in the ‘tty/<pid>’ which can be used to interact with it remotely - much wow

e.g. $ socat - UNIX-CONNECT:tty/38284

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

This reverts commit 91a2945e.

Conflicts:
	daemon/worker.h

this avoids repetitive allocation/free cycles on load

the engine keeps an array of storage backends available
for caching, it is possible to switch the backend on
runtime with standard ‘proto://config' string

the default backend is ‘lmdb://<path>’

refs #15

lists available interfaces

The engine is responsible for central managemento of resources and work
distribution. This allows to spawn multiple loops.

The engine configuration / interface is based on Lua 5.1-5.2, including
a RPC and a simple CLI interface. This is going to be the core of configuration
later on.

- cache is now fully optional, resolution works without cache as well
- proper mempool per each request, not mixed
- zone cut resolution doesn't need any per-context memory
- now it's possible to do qname minimization
- rplan keeps tabs on resolved queries
- added documentation etc.