SWE-Perf: Can Language Models Optimize Code Performance on Real-World Repositories?