Leveraging Trajectory Optimization To Improve Deep Reinforcement Learning, With Application To Agile Wheeled Robot Locomotion