Why does fine-tuning a pretrained language model (like BERT) often outperform training a task-specific model from scratch on a smaller labeled dataset?