Clarify limitations of LanguageModelFeaturizer in docs

[mleimeister]

This docs update aims to clarify some details around which HuggingFace models can be used with LanguageModelFeaturizer. Despite the supported architectures being listed in our docs, users often try others and report errors in the forum. Furthermore, not all pretrained weights from the HF models hub for the supported architectures can be used (e.g. due to missing TF weights or non-standard tokenizers being used), which results in error messages that are hard to interpret.

A related ticket with links to forum issues can be found here here.

Proposed changes:

• Clarify the limitations of the current implementation of LanguageModelFeaturizer w.r.t. model architectures and weights from HuggingFace

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)

[github-actions]

🚀 A preview of the docs have been deployed at the following URL: https://10616--rasahq-docs-rasa-v2.netlify.app/docs/rasa

[mleimeister]

@koernerfelicia @dakshvar22 Hi, I just wanted to check what you think and if this would improve the LM featurizer docs, since we've had frequent questions about this on the user forum recently. Not sure if I'm missing anything on the requirements for model weights to be useable with our implementation, but those listed were the ones I experienced as causes for errors.

[koaning]

+1 for picking this one up @mleimeister! I've added a few comments.

[koaning]

This looks grand!