No Strong Feelings One Way or Another: Re-operationalizing Neutrality in Natural Language Inference