Towards Decoding as Continuous Optimization in Neural Machine Translation