Train-test split in training and inference set in forecasting

[mariamedp]

Main changes:

• Split original dataset by date (same as AutoML) and change references of second set to inference set (prev. test set)
• Train-test split during training. Metrics are computed on test. Extra param "test_proportion" that defaults to 0.3.
• After logging the metrics, model is retrained using the whole dataset (training part, always leaving aside inference set) and registered
• Forecasting script no longer calculates and logs metrics, it only generates predictions (as in AutoML)