Instruct Once, Chat Consistently in Multiple Rounds: An Efficient Tuning Framework for Dialogue