UniGuard: Towards Universal Safety Guardrails for Jailbreak Attacks on Multimodal Large Language Models