Learning Quadruped Locomotion Using Differentiable Simulation