apple-aiml-research /ml-rsi-inference-benchPython1347A Long-Horizon Benchmark for Self-Improving Inference Systems