Optimality of affine policies in multi-stage robust optimization