SWE-Bench: Can Language Models Resolve Real-World GitHub Issues?