Add self attention encoder

[Adam]

With the decoupling of encoders and decoders, we have added a Linear encoder, which seems to just embed the inputs and pass them along. We should also add a SelfAttention encoder, which encodes the embeddings with a self attention layer (and no positional encoding).

This contextualizes the embeddings by representing each as a linear combination of itself wrt all other embeddings.

[Kyle Gorman]

+1. Makes sense.