Consistency-based Abductive Reasoning over Perceptual Errors of Multiple Pre-trained Models in Novel Environments