Think, Prune, Train, Improve: Scaling Reasoning without Scaling Models