TOFD: Target-Oriented Feature Decoupling against Poisoning Attacks in Split Federated Learning