Boosting Graph Neural Networks via Adaptive Knowledge Distillation