Remove FeatureApplicator and use DataPreparation and FeatureAdders instead

[BenoitHardier]

This issue is a medium-long term issue related to the data preparation phase (before fit / predict methods).
In several previous issues, we introduced some workhorses to deal with features and data preparation in a flexible way, using the python capacity to link dynamically with custom code.

• In [Feature] Review/give feedback to FeatureAdder implementation #187, we introduced the feature_adder and feature_dispatcher objects used to add new features.
• In 434 autoregressive model and more generic feature applicator #437, we introduced the DataPreparation class that should handle all the data preparation computation.

For now, we keept the legacy way to build feature thanks to FeatureApplicator classes and implement very naively the LegacyDataPreparation class that wraps the FeatureApplicator calls.

We think that:

1. The 2 FeatureApplicator (train and Operational) can be entirely replace by one DataPrepartion class with a method for each task (train and forecast)
2. The LegacyDataPreparation can be be directly implemented without the use of the FeatureApplicator
3. Most of weather and lag feature, historically in the library for alliender's usecases, can be implemented through the feature_adder mecanism to completely get ride of specific hardcoded parts.