Support logging of accuracy for non-constant answer choices

[AlexTMallen]

Huggingface evaluation doesn't allow passing of the answer choices to the evaluation metric function, so we are currently asserting that the answer choices for all examples are the same.