Language Models from Transformers Lib

[dakshvar22]

Proposed changes:

• Create a new NLP component - HFTransformersNLP which tokenizes and featurizes incoming messages using the Transformers Library.
• Create LanguageModelTokenizers and LanguageModelFeaturizers which use the information from HFTransformersNLP and sets them correctly for message object
• Architectures supported: Bert, OpenAIGPT, GPT-2, XLNet, DistilBert, Roberta

Part of https://github.com/RasaHQ/research/issues/62

Status (please check what you already did):

• added some tests for the functionality
• updated the documentation
• updated the changelog (please check changelog for instructions)
• reformat files using black (please check Readme for instructions)

[Ghostvv]

Looks good! Have a couple of comments. Also, does it makes sense to create 3 files for hf? Why don't we put all these helpers into 1 file?

[tabergma]

We need to add some tests for the components.

[tabergma]

Can you also create a changelog entry and add some documentation? E.g. add the new components to https://rasa.com/docs/rasa/nlu/components/.

[dakshvar22]

@tabergma Yes, tests and documentation are to be added. That wasn't ready. :)
@Ghostvv I feel readability is better. From a maintenance perspective, we know that registry.py and transformers_pre_post_processors.py is where bulk of it would happen since new models can basically be added by editing those two files. IMO the three files help in that sense. What do you think?

[Ghostvv]

Are we going to add when to use what in a separate PR?

[tabergma]

Apart from the comments I already made, it looks good 🚀 Great work!