the server responds with edns0 if the client requests it,
it also uses DNSSEC for queries if DO=1

the obtained records are not however validated

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>’

the api is still a global variable, but it would be the
best if the namedb_t carried the pointer to the api,
so it needs to change in libknot later

each asset is tagged by a byte defining its type and importance
this is a groundwork for negative cache and packet cache
abstracted the code in preparation for different backends

do all socket closing in worker to avoid double closes with
timeout timer, also propagate bad messages to discover errors
earlier than timeout

if the server doesn’t understand edns0, replies with
notimpl, formerr or sends bad badly formed response,
we try basic dns query (ns0.rbsov.bbc.co.uk)

previously cryptolib random function was used to generate message id,
this works well but it is slow especially when the entropy is low,
replaced with cryptographically safe prng ISAAC

the ccan directory is going to be used in the future, as it's include
structure makes it easy to embed C snippets instead of reimplementing
them

the encloser might exist and contain valid data, the resolver would return wrong records in this case

refs #15

this disables dual-stack and allows binding to both v4 and v4-in-v6
addresses separately

fixes #22

* each query is assigned a task
* each task contains request, some primitives and mempool
* worker can process multiple tasks at once and
  offload I/O to event loop

Not finished:

* it depends on icmp/system timeouts, #22
* tcp reads are going to be bad if the messages
  arrive fragmented #21

the reason is that LMDB allows only 1 write transaction
per thread, and this doesn’t work well with multiplexed
requests. there is probably going to be some performance hit,
but now transactions are scoped for each cache operation

the worker now creates a resolution context copy,
and keeps it if the query requires iterative queries.
the worker_exec() is now a reentrant function that gets
called with incoming data until the resolution is done,
and it sends the answer

TODO: per-loop freelists are implemented