Refactor meta-learning components to single data transformation

[MorrisNein]

• Some current meta-learning components include multiple transformations with reuse of other ones.
• Moreover, there is excess usage of hardcoded configuration and data files affecting the components' behaviour.
• The usage of cache is ambiguous. The interface of the components must clearly involve cache retrieval and corresponding notification.
• The current cache implementation may be reconsidered to become a universal lazy computational cache.

These major problems affect a big part of the framework, thus must be solved step-by-step:

• Clean the component SurrogateAdvisor #73
• Reconsider the interface of PymfeExtractor #75
• Clarify cache usage  #77
• Unify the input interface of MetaFeaturesExtractor #76