Less is More: On the Feature Redundancy of Pretrained Models When Transferring to Few-shot Tasks