A Model-Agnostic Approach to Differentially Private Topic Mining