Relaxing Continuous Constraints of Equivariant Graph Neural Networks for Physical Dynamics Learning