Long Is More Important Than Difficult for Training Reasoning Models