Meta-Learning with Versatile Loss Geometries for Fast Adaptation Using Mirror Descent