On the Power of Small-size Graph Neural Networks for Linear Programming