Logprobs eval

[chanind]

This PR takes some of the ideas from in #56 and formalizes them in our architecture with tests to make this sort of eval easy moving forward. Specifically, this adds the following:

• Adds the awesome make_dataset() changes from @dtch1997 and adds test coverage around the splitting behavior
• Adds a Pipeline.calculate_output_logprobs() method based on the code in the the Jupyter notebook in Truthful QA benchmark #56, and adds test coverage.
• Adds a MultipleChoiceAccuracyEvaluator which implements the accuracy calculation from the Jupyter notebook using logprobs within our Benchmark framework, including test coverage.
• Adds the hardcoded TQA data from the notebook.

This PR also changes our EvalPrediction and Evaluator types to support logprobs. Now, each Evaluator must specify if it requires_generation or requires_probs to indicate to the benchmark what needs to be run. The benchmark will run generation and/or calculate probabilties as required by evaluators.

I also moved the make_dataset() stuff from data/__init__.py into data/make_dataset.py to make it easier to test.

[dtch1997]

Overall LGTM! Great work implementing the MCQ-style log prob evaluation as well as expanding the test coverage