Update forecast loop

[capelastegui]

At the moment, run_forecast takes candidate trend and seasonality models, and fits all possible combinations. This gets very inefficient as new models are introduced. We should move towards a stepwise flow, where models are added incrementally, and only the best-fitting models progress to subsequent steps.

New workflow:

• normalize_df_in(df_in_raw): Normalize input. Returns df_in (normalized) and df_norm_map: table that can be used to restore column names
• get_df_models(df_in): Generate table with models to use, in sequence, to forecast each source. Returns df_model
• forecast_fit_single(df_in, source_id, df_models): Fit forecasts for single source, sequentially. Returns df_metadata
• get_df_forecast(df_in, df_metadata, t_extrapolate, add_pi, include_all_models): Extrapolate forecast for models. Returns df_forecast
• denormalize_df(df_forecast, df_norm_map)

Notes:

• season_add_mult - relevant in forecast_fit_single, not get_df_models