Tactile-VLA: Unlocking Vision-Language-Action Model's Physical Knowledge for Tactile Generalization