Fix str(res) typo in post_check for non-numeric answers

[rrrxxx0510]

Fixes #1679

The non-numeric branch of post_check compared str(res) where res is always None, so free-form questions with a non integer/float answer type could never be scored correct, even when the GPT judge extracted the right answer.

Day-one typo from #29. lens.py and ocr_reasoning.py, which were derived from the same template, correctly use str(response) here — confirming the intended behavior. qspatial.py copied the same typo; fixed both.

On the shipped MathVista_MINI TSV the blast radius is the 2 samples with answer_type == 'list' (the 540 text answers are all multi-choice and take the other branch), so this mostly matters for correctness and for any future/custom TSVs with free-form text answers.

Verified with a direct call into post_check before and after: a correct text answer was marked wrong before, and is scored correctly after; wrong answers and integer answers behave unchanged.

[ssiq]

Thanks for the updates. These three commits seem to address separate issues:

1. preserving structured extra_records
2. adding a Transformers model-class fallback
3. fixing the str(res) typo in MathVista/QSpatial

Would you mind splitting them into three separate PRs? That would make the changes easier to review, test, and merge independently.

[rrrxxx0510]

Both PRs are now single-commit — the fallback fix stays in #1676 and the typo fix in #1680. The extra_records change was never supposed to be in there, it sits on its own branch and I'll send it as a separate PR once #1665 gets some feedback. My bad for stacking the branches, and thanks for catching it.

[rrrxxx0510]

按你说的拆好了，单 commit——typo 这条单独留着，fallback 那个归 #1676。

force-push 之后 CI 批准被重置了，两个 check 还都是 action_required 没跑起来。麻烦再看下 workflow？