Why Stable Learning Works? A Theory of Covariate Shift Generalization