The Why And How Of Label Variation In Natural Language Inference