Orthant-Wise Passive Descent Algorithms for Training L1-Regularized Models