Adding Encoder-decoder attention cache. https://github.com/tensorflow/tensor2tensor/pull/827 - View it on GitHub