Scalable Best-of-N Selection for Large Language Models via Self-Certainty