CACTO-SL: Using Sobolev Learning to improve Continuous Actor-Critic with Trajectory Optimization