Make Domain accessible in NLU Pipeline

[dakshvar22]

Description of Problem:
Currently the domain.yml file is read and made accessible only in the Core pipeline. As we include more features into the complete stack, there may be some which do not follow the clear boundary between NLU and Core. One such example feature is Supervised Response Selector. The particular part where the domain file is needed is described in this comment .
There may be more such use cases where this might be necessary.

Overview of the Solution:

• We'll need to move out domain.py outside of rasa.core since it would be shared between NLU and Core.
• For the training phase, similar to how domain is loaded inside _train_core_with_validated_data() function we can load it inside train_nlu_with_validated_data() and pass it to rasa.nlu.train() function.
• Both Trainer and TrainingData could potentially need access to the Domain instance.
• During inference, individual components may need to use the domain data for processing the message.

Blockers (if relevant):
Any ideas?

Definition of Done:

• Tests are added
• Feature described the docs
• Feature mentioned in the changlog
• Compatibility test for existing features

[msamogh]

You'd also need to ensure the fingerprints don't step on each other, because right now, the whole domain just gets one single fingerprint.

[dakshvar22]

Not pursuing this for now, as the primary use case was solved in a different way #4233