Cybench: A Framework for Evaluating Cybersecurity Capabilities and Risk of Language Models