$L_2$-Regularized Empirical Risk Minimization Guarantees Small Smooth Calibration Error