DeepConf: Scaling LLM reasoning with confidence, not just compute