Feature/guidedalignment

[jungikim]

This request adds the implementation of the guided alignment as described in Guided Alignment Training for Topic-Aware Neural Machine Translation (Chen et al. 2016)

In summary:

• preprocess.py: optionally stores source-to-target alignments in parse format (-alignfile, -alignvalfile)
• s2sa/data.lua: if present, loads the alignments and converts them into dense format per batch
• train.py: optionally creates a parallel criterion consisting of the decoder criterion (ClassNLLCriterion) with the criterion for guided alignment (MSECriterion) (--guided_alignment, --guided_alignment_weight, --guided_alignment_decay)
• s2sa/models.lua: optionally exposes attention output in the decoder model

[yoonkim]

Cool! Also relevant: http://arxiv.org/pdf/1609.04186.pdf (the above work claims that cross entropy does slightly better than MSE for training the attention part of the model)