model_builder adaptations for full sklearn compat

[michaelraczycki]

Adaptations for full sklearn compatibility:
All functions expect data to be passed around in X,y format known from sklearn
output_var now is a property, which allows to use it in loading functions (still save saves entire original dataset - concatenated X and y, so this was needed to be able to always tell which column is a target column)
fit now requires X to be pd.DataFrame and y to be pd.Series

[michaelraczycki]

also in order to simplify the code, now the model_config and sampler_config can be passed only at the class initialization, other approaches where they could be passed to build model required constant checks and was polluting the code, without clear benefits.

[twiecki]

Can we make y be optional.

[michaelraczycki]

sure, it will introduce some complications in child classes, as then we're technically allowing to build with y being None, and DelayedSaturatedMMM is build assuming that it's there. In general it will add probably quite few if's in the code but it shouldn't be a big deal

[twiecki]

The issue is that if we don't do that it will limit the models we can fit this on, as not all actually have an X and a y, for example, the change-point model only has a single time-series.