Bigger is not Always Better: Scaling Properties of Latent Diffusion Models