Learning Neural Constitutive Laws From Motion Observations for Generalizable PDE Dynamics