Improving Weak-to-Strong Generalization with Reliability-Aware Alignment