Fixing `LitQAEvaluation` bugs: incorrect reward indices, not using LLM's native knowledge

This PR:

Fixes both issues identified in https://github.com/Future-House/paper-qa/issues/693
1. Using LLM's innate knowledge, which this PR fixes by updating EVAL_PROMPT_TEMPLATE
2. Incorrectly handling options beyond the specified options, which this PR fixes by declaring those as "incorrect"
Fixes bad discounted_returns logic and incorrect reward indices
- Previously, incorrect answers were given 0.1 reward, and unsure answers were given -1.0 reward
- Adds test coverage of this, both upon discounted_returns and the TaskDataset
Makes the input distractors be a Sequence to avoid in-place edits, further robust-ification after https://github.com/Future-House/paper-qa/pull/694

Future-House / paper-qa