Linked Presentation: Tight Auditing of Differentially Private Machine LearningExtracting Training Data from Diffusion Models