Distributed NLI: Learning to Predict Human Opinion Distributions for Language Reasoning