We should create an `INLUBatchTestClient` interface for NLU providers that accept a batch of utterances as opposed to individual prediction requests.