See What the Robot Can't See: Learning Cooperative Perception for Visual Navigation