To Repeat or Not To Repeat: Insights from Scaling LLM under Token-Crisis