g8a9/l3wrapper: Add parallelization and rule sets modifiers.

This version adds the support to parallelization, e.g. L3Classifier can be used with multiprocessing or as an estimator in scikit-learn's GridSearchCV with (n_jobs=-1).

Additionally, there is the first release of rule set modifiers, which enable changing the rule sets after the training step. Currently "level1" is supported, i.e. L3Classifier will keep the level 1 rules only and discard level 2 right after training. Please note that a model instantiated this way and then saved, will have an empty lvl2_rules_ parameter list.