Referee-Meta-Learning for Fast Adaptation of Locational Fairness