Revisiting k-NN for Fine-tuning Pre-trained Language Models