* Move recipe processing code in the worker;
* Refactor CLI and use the the worker;
* Implement Recipe#tasks and remove tasks application during evaluation,
tasks are now applied by the worker after all evaluations are done.
* Report invalid action usages from tasks;
* Implement backtrace cleaning in CLI;
* Extract error class declarations in a new errors file;
* Replace raise with fail keyword in task DSL class.
Currently the idea is to later implement a -h (or -t) option to the
command, and/or a `target' recipe keyword. Another option might be to
use the target/host key when the key/value registry will be implemented
and made available to the recipe and task DSLs.
