On Coresets for Regularized Loss Minimization