Deep Declarative Dynamic Time Warping for End-to-End Learning of Alignment Paths