IMPLI: Investigating NLI Models’ Performance on Figurative Language