Exploiting the Potential of Seq2Seq Models as Robust Few-Shot Learners