DCNN Layers

[gregdhill]

The new model, previously introduced by Nal Kalchbrenner, has been implemented in PyTorch, but due to time constraints for the original publication I implemented the k-max pooling and folding operations in Python. To improve the speed of training it would be better if these were integrated with PyTorch at a lower-level.