Add SentenceTransformer to NLQ

[miguelmartin75]

Caveat: since SentenceTransformer reduces the input features from [B, T] -> [B, 768], I don't pass the token embeddings to the shared feature_encoder

[satwikkottur]

Thanks @miguelmartin75 for this PR.

Two comments:

1. Can we have choices for --predictor flag in options.py? Since we're adding addition predictors, would be nice to restrict that set.
2. Between st and sentence_transformer (or even sent_transformer), I prefer the latter for readability. Thoughts?