Enabling passing context into `Model.forecast`

[Mr-Geekman]

🚀 Feature Request

We decided that we want to enable passing context into forecast method of the model. It will be done with a new parameter: prediction_size. First n - prediction_size points in a given ts won't be forecasted and if model needs a context, it will use this points for this.

Proposal

1. Add prediction_size parameter for the adapters.
   • implement separately common logic on selection of last prediction_size elements of prediction
2. Add prediction_size parameter to forecast method in classes:
   • base.Model
   • base.ForecastAbstractModel
   • base.PredictIntervalAbstractModel
   • base.PerSegmentModel -- determine the context needed for prediction and pass it into the adapter in df_context parameter, if context isn't needed don't pass df_context.
   • base.PerSegmentPredictionIntervalModel -- same as above
   • base.MultiSegmentModel -- same as above
   • base.DeepBaseAbstractModel -- rename horizon to prediction_size
   • base.DeepBaseModel -- rename horizon to prediction_size
   • add to other classes if it wasn't enough
   • default value should be None and in the body of the function correct value is determined as a maximum possible prediction_size
   • if a too big prediction_size is set (not enough context) throw an error
3. Fix forecast in pipelines to work with context
   • Fix Pipeline.forecast, the code should be similar to snippet with DeepBaseModel
   • Fix AutoRegressivePipeline.forecast
   • Fix ensembles.

Test cases

1. Make sure current tests pass.
2. Test function on selection of last prediction_size elements for adapters.
3. Test that if prediction_size is None then value is determined as len(df) - context_size. Probably, it is convenient to do with some dummy/mock model.
4. Test that models throw error if prediction_size is too big. Look above about dummy/mock.
5. Test that Pipeline.forecast passes big enough dataset according to `horizon. Look above about dummy/mock.
6. Test that AutoRegressivePipeline.forecast passes big enough dataset according to horizon. Look above about dummy/mock.

Additional context

• Is blocked by #844.
• Merge it into an inference branch.