More encodings of LMs

[V-G-spec]

After recent conversation with Yifan, we need the following encodings for all the models (or for as many as possible): The one after the embedding layer The one after the encoder bit of the model (For the few models which don't have the encoder-decoder architecture, anything from the middle layers would work) The final representation (If I understand correctly, this is what @noah already wrote the script for)

[tamas-visy]

More specifically, we could use any numerical representation from the model, so any embedding between the initial embedding and the final representation could be investigated.