Leveraging Pretrained Latent Representations for Few-Shot Imitation Learning on a Dexterous Robotic Hand