Cut-and-stitch: efficient parallel learning of linear dynamical systems on smps