A Critical Review on the Use (and Misuse) of Differential Privacy in Machine Learning