Unleashing the Power of Randomization in Auditing Differentially Private ML