Integrate rasa train and rasa run with new model architecture

[wochinge]

Description
It's time for a big step 🔥 After having migrated all components, implementing rasa.engine, creating the new components for the graph, and implementing the graph generation using the recipes it is time to integrate everything together.

Definition of Done

• rasa train trains using the graph architecture (+ the variants like rasa train nlu/core)
• rasa run loads graph architecture models and is able to make predictions with it
• The test suit passes
• the model accuracy didn't drop (I think we need to do that as part of this one as it will be a pain to find the issue later on; also not expecting any issues here)

Things which don't need to be done

• Making sure all no longer needed code is deleted (we can do that separately)
• Assessing / improving performance will be a separate issue

[wochinge]

• fix caching of empty resource
• fix recipe tests
• not load predict thing as part of train
• rasa train @wochinge
• rasa predict @joejuzl (maybe: HTTPInterpreter 😁 )
• fix recipes tests @wochinge
• fix check logging @joejuzl

Ideas:

• targets for prediction should be in metadata
• put cli + language in execution context for train