Universal Few-shot Learning of Dense Prediction Tasks with Visual Token Matching