Clean up after TF 2 and new NLU architecture

[tabergma]

Clean up/Tackle remaining open tasks from #5266

• Add docstrings and types to all methods, classes (especially in policies and NLU components)
• Split up long methods in DIETClassifier (e.g. _create_model_data, _predict_label) and in hf_transformers.py(e.g._load_model`)
• In DIETClassifier: Throw warning or error on not enough labels? (should be consistent with other classifiers, e.g. SklearnIntentClassifier)
• Check if the spacy version check in SpacyTokenizer is still needed
• Consider moving the tensorflow package out of the utils package

[m-vdb]

@ricwo is there anything to discuss here or should we move it to backlog?

[Ghostvv]

@samsucik I feel like the models part is addressed by your PR: https://github.com/RasaHQ/rasa/pull/7589/files/0aaf2646bed3b515c50ebb82e716e6c1450c3f3c..6003daee3941229ec56ecb545d7ce9a7a789532e

@dakshvar22 is anything additional needs to be done for hf transformers?

things like checking spacy version and moving packages around are engineering tasks

[samsucik]

@Ghostvv just by reading the issue, I think my PR addresses the models part only partly, not fully...