Add Prompt Tuning

[lenglaender]

Depends on: #591

This PR adds Prompt Tuning (https://aclanthology.org/2021.emnlp-main.243/)

Currently the Prompt Tuning Layer was created and the BERT model has been adapted

[calpt]

did a first review pass, looks good overall, haven't checked all details though

[calpt]

Please also add row in table on main readme

[raphaelreimann]

Hey, love the integration of prompt tuning! I would love to use it when it is stable, do you know when this will be merged? 🙂 @lenglaender @calpt

[lenglaender]

@raphaelreimann We're currently fixing the last few tests that aren't running. For every model except CLIP, EncoderDecoder, GPT-2, GPT-J, LLaMA and T5 the implementation is already working.

As soon as this PR and the sync PR (#602) are merged, we will release the new Adapters library (#584).

[raphaelreimann]

@lenglaender Thank you for the update on this, great work!

[hSterz]

Looks good to me just a few minor comments

[lenglaender]

For the subsequent PR, which will extend Prompt Tuning with composition support and which will add support for the not yet supported models, we should think about realising the change of the attention mask for prefix tuning within the function prefix_attention_mask.