Self-Rationalizing Models for Commonsense Reasoning