Low-rank lottery tickets: finding efficient low-rank neural networks via matrix differential equations