SpeechGuard: Exploring the Adversarial Robustness of Multimodal Large Language Models