Optimal Transport for Structure Learning Under Missing Data