PGCSPose: Physics-Constrained Generation and Causal Semantic Fusion for Robust In-Hand Pose Estimation