Diagnosing the Compositional Knowledge of Vision Language Models from a Game-Theoretic View | Synapse