Learning Quadrupedal Locomotion via Differentiable Simulation