Revisiting regression adjustment in experiments with heterogeneous treatment effects