Reversible Instance Normalization for `RegressionModel`

[dennisbader]

Is your feature request related to a current problem? Please describe. Reversible Instance Normalization (RIN) has proven to be very effective for our neural network models (against distribution shift, ..). We could also add this to RegressionModel. I see potential especially for tree-based models, where RIN could help for data subjected to trend or increasing variance.

Describe proposed solution

• Add use_reversible_instance_norm to RegressionModel
• Transform the target lagged features in the X, and y arrays during training
• Inverse transform y_pred