HumanAgencyBench: Scalable Evaluation of Human Agency Support in AI Assistants