The Sensitivity of Language Models and Humans to Winograd Schema Perturbations