Lightweight Diffusion Models with Distillation-Based Block Neural Architecture Search