Toward a Human-Centered Evaluation Framework for Trustworthy LLM-Powered GUI Agents