Learning Fast Samplers for Diffusion Models by Differentiating Through Sample Quality