Modeling Long Context for Task-Oriented Dialogue State Generation