RL on Incorrect Synthetic Data Scales the Efficiency of LLM Math Reasoning by Eight-Fold