VHELM: A Holistic Evaluation of Vision Language Models Chi Heem Wong