Towards Irreversible Machine Unlearning for Diffusion Models