Rethinking the Sample Relations for Few-Shot Classification