Accelerating Visual-Policy Learning through Parallel Differentiable Simulation