Make entity extraction for entities separated by comma configurable

[akelad]

With PR #6191, entities separated by comma now get extracted separately. This makes sense for something like:

please add buffalo, ranch, mustard, barbeque sauces results in {'entity': 'sauces', 'start': 11, 'end': 43, 'value': 'buffalo, ranch, mustard, barbeque', 'extractor': 'DIETClassifier'}

But maybe shouldn't work this way for address entities like 5306 Walnut Ave., Building A, Sacramento, CA 95841. See also issue #6795.

We should make this behaviour configurable in the entity extractors, and maybe even on an entity level, given you might have both types of entities in one bot.

[tttthomasssss]

@tabergma @akelad just to be clear - is the idea to make this configurable per entity type (on the basis of the entity types that are annotated in the NLU training data)? For example, supposing that address and ingredients are two entity types in the data, users simply need to specify sth like the following in the config (with default behaviour of True?):

language: en
pipeline:
  ...
  - name: DIETClassifier
    epochs: 100
    entity_recognition: True
    split_entities_by_comma: 
      address: False
      ingredients: True
  ...

[tabergma]

Something like this sounds good I would say. I guess it would be even better if we support both variants:

language: en
pipeline:
  ...
  - name: DIETClassifier
    epochs: 100
    entity_recognition: True
    split_entities_by_comma: False/True
  ...

and

language: en
pipeline:
  ...
  - name: DIETClassifier
    epochs: 100
    entity_recognition: True
    split_entities_by_comma: 
      address: False
      ingredients: True
  ...

We also should be able to handle the cases where not all entities are listed in the second configuration. And this should also be available for the CRFEntityExtractor.