A pull request for #242 (greedy decoding and vectorization in attention.py)

[emanjavacas]

This addresses both issues. I've kept comments with tensor dimensionality in the computation of the attention weights, just tell me if I you'd prefer them removed.

[neubig]

This is great, but w1dt = w1 * input_mat can be done only once at the beginning of the sentence and cached. This is a big performance win, so it'd be nice to add that as well (maybe with a comment).

[emanjavacas]

Oh, you are right. Here is the new version.

[neubig]

Looks good! Thanks.