Unravelling the Performance of Physics-informed Graph Neural Networks for Dynamical Systems