Learning Visual Locomotion With Cross-Modal Supervision