Do Differentiable Simulators Give Better Policy Gradients?