[Feat] Scaled Distributions

[kdgutier]

This PR has no implications for point forecasting methods.
The PR incorporates a reinterpretation of DeepAR's training strategy that scales the DistributionLoss.

• Added identity possibility to TemporalScaler class, to simplify train/validation/predict_steps from Base classes
• Simplified code with identity scaler, eliminated if scaler is None conditions in train/validation/predict_steps.
• Changed BaseRecurrent and BaseWindows train/validation/predict_steps to accept Scaled distribution.
• Changed model notebooks to perform qualitative tests of the DistributionLoss.

PENDING:

• Add mean_scaler with shift=0 and scale=mean(x) to Temporal scalers.
• Add outsample_y recovery unit test with _inv_normalization method (Poisson non-negativity failing).
• Change names from loc -> shift to homogeneize scalers and Scaled Distributions.
• Change back scaler_type default to None.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB