Simplicity Bias of Transformers to Learn Low Sensitivity Functions