Differentially Private Coordinate Descent for Composite Empirical Risk Minimization