Towards Safe Self-Distillation of Internet-Scale Text-to-Image Diffusion Models