Deprecate training_interval config item

[dominiquekleeven]

The training_interval (or fitting interval) of the model is currently configurable.
This provides no real benefits and can potentially cause issues when generating forecasts.

Explanation

Prophet and other time-series forecasting models generate their forecasts starting from the last known training (fitting) datapoint.
If the model is retrained every 2 hours but forecasts are generated every hour, the resulting forecast will miss 1 hour of future data.

The model should either:

• Always train immediately before forecasting, or
• Clearly define and enforce a consistent relationship between training and forecasting intervals.

Task

• Deprecate the training_interval config item and replace its usage with forecast_interval, resulting in both the model training/fitting and forecasting to be aligned.

Notes:

• The system's task scheduler enforces training/fitting to be performed before forecasting if their execution times overlap, so this is already taken care of.
• Fitting/training time of the model is negligible, Data processing improvements #43 is a work in progress PR that also improves speeds substantially.