PhyVLLM: Physics-Guided Video Language Model with Motion-Appearance Disentanglement