Identifying and Disentangling Spurious Features in Pretrained Image Representations