PaperBench: Evaluating AI’s Ability to Replicate AI Research