fix case of special tokens in encoder

[benlipkin]

special tokens, e.g. , from tokenizer cause 1-off errors when using indices to extract sentence representations from context.

[aalok-sathe]

Now what happens here is: the special tokens are chopped off from each stimulus when extracting stimulus-level representations evaluated within a context.
The remaining thing here is: being able to extract first-token/last-token/special-token representation for a single stimulus, because now special tokens are chopped off by default since in context they represent the whole context rather than any stimulus

[aalok-sathe]

whoops, that was an incorrect reference to this issue. it should have been #18 instead