Move input generation and inference for each class of models into distinct modules

[hvgazula]

The input generation, inference, and embeddings/logits extraction functions (as appropriate) tfsemb_main.py should be moved into separate scripts for causal, mlm, and seq2seq models.

[zkokaja]

Yes, let's flesh this out and implement a prototype for causal so we can implement it for other types as well.

[zkokaja]

Consider adding special tokens to causal models at the beginning, to stay true to what the model was trained on. Needs investigating.

[hvgazula]

https://datascience.stackexchange.com/questions/86566/whats-the-right-input-for-gpt-2-in-nlp

[zkokaja]

Let's talk about this again

[zkokaja]

waiting to resolve issue with whisper embedding generation replication