Learning Integrable Dynamics with Action-Angle Networks