Stress testing

[kasnerz]

We should test as many aspects of the app as we can, as thoroughly as we can.

That means:

• [x] adding new datasets of various types (see #20, #21 )
• [x] running LLM evals with various configs (see #20, #24 )
• [x] running human evals with custom templates (at least internally, perhaps even on Prolific?) (See #20 )