Learning an Efficient Multi-Turn Dialogue Evaluator from Multiple Judges