Model Uncertainty-Aware Knowledge Amalgamation for Pre-Trained Language Models