Helpful or Harmful? Evaluating LLM-Assisted Vulnerability Patching via a Human Study