Differentially Private Sampling from Rashomon Sets, and the Universality of Langevin Diffusion for Convex Optimization