Efficient Fully-Offline Meta-Reinforcement Learning via Distance Metric Learning and Behavior Regularization