On the relationship between empirical likelihood and empirical saddlepoint approximation for multivariate M-estimators