A New Perspective on Boosting in Linear Regression via Subgradient Optimization and Relatives