RedStar: Does Scaling Long-CoT Data Unlock Better Slow-Reasoning Systems?