Revisit Few-shot Intent Classification with PLMs: Direct Fine-tuning vs. Continual Pre-training