Meta-learning the mirror map in policy mirror descent