Accelerated Training for Matrix-norm Regularization: A Boosting Approach