Toward Understanding the Transferability of Adversarial Suffixes in Large Language Models