Test that LM queries are being made correctly

[kcarnold]

• are logprobs being computed from the results correctly?
• is spacing between sentences sensible / consistent during reorderings?
• max_tokens should be set tiny since we're not actually using the generation results
• could there be better prompts for the LM (vs the "write a short essay using this outline" that's there)

It would be helpful to have some test cases that didn't need an OpenAI API call (just paste in a prior response).

It would be helpful to visualize what's going on. For this issue, at least log the LM queries in a way that we can show easily. In general we probably want a more visual representation of what's going on; full version of this is a separate issue.