Implement Recipe to convert current config to GraphSchema

[wochinge]

Overview of the Solution:
Implement the Recipe class which converts the current config to a train and predict GraphSchema using the given CLI params. This component loads all the classes and uses the @DefaultV1.register(...) decorator to create the graph structure from it.

See the implementation proposal for further details.

Blockers (if relevant):
Existing components needs to be adapted to new graph interface.

Definition of Done:

• Recipe is tested with different configurations (default configuration, configurations with custom components / policies, auto configuration)
• Recipe interface is implemented
• all default Rasa Open Source components / policies are decorated with the @DefaultV1.register(...)
• the configuration is validated
  • Check pipeline is not empty
  • Checks only one tokenizer is present

[wochinge]

We could consider omitting Duckling (and other pretrained extractors) from the train_graph 🤔 See #3389 (comment) .

[wochinge]

omit e2e featurization when we only train a core model

[m-vdb]

@wochinge and @joejuzl to split this issue

[wochinge]

Moved this back to Backlog as the core functionality is implemented but now have to until all things have been changed to be GraphComponents before wrapping this up