The Dark Side of Function Calling: Pathways to Jailbreaking Large Language Models