Enhancing Robustness and Efficiency of Least Square Twin SVM via Granular Computing