One Model, Any Scenario: End-to-End Locomotion from Vision