Learning Physical Dynamics with Subequivariant Graph Neural Networks