Extend pipeline functionality to classical forecasters

[mloning]

• See pmdarima Python package for a working example
• related to TSC/TSR: implement pipeline building functionality with transformers on y #7 and Design/implement transformers on y #44

[mloning]

Useful transformers on target (y) include

• detrending
• deseasonalizing

Transformations have to be applied before passing to prediction model and inverse transformations after prediction model (i.e. on predicted values)

[mloning]

Following #44 and scikit-learns TransformedTargetRegressor, I would implement a TransformedTargetForecaster to handle y-transformations.

But in a time series setting,

• there may be some y-transformation where one would not want to do an inverse transform on y after prediction, e.g. smoothing transforms of y, not sure how to handle this?
• one may want to chain multiple transformations (i.e. in a pipeline), but passing a pipeline of transformers would not work, as currently sklearn's Pipeline does not have an inverse_transform method.

[mloning]

Closed by #111