Carry a mixed document's layout rules on ParseTestCase

[boyang-zhang1]

Builds on #133 — that PR's commit by @AmirF194 is the first commit here, so its diagnosis, fix and regression test are preserved. Stacked on top of #153, which hardens the evaluation path these documents now reach.

#133 correctly identified that _load_jsonl_dataset dropped the layout rules of a document that also carries parse rules, while the run still reported success: true. It routed those documents to LayoutDetectionTestCase. This commit keeps them on ParseTestCase instead, because that class already type-routes layout, extract_field and parse rules on one test_rules list and is the only branch that carries the parse-side state.

Why the container matters

	LayoutDetectionTestCase	ParseTestCase
expected_markdown	no such field	carried
allow_splitting_ambiguous_merged_tables, trm_unsupported, max_top_title_rows	no such fields	carried
test_rules type	closed union of the built-in rule models	SerializeAsAny[ParseRuleBase]

Three concrete consequences, all reproduced against main + #133:

1. Markdown and table settings vanish. Same JSONL document (layout rule + table rule with settings + inline expected_markdown): ParseTestCase | md='| a |' trm=True split=True maxtop=3 becomes LayoutDetectionTestCase | md=<no field> trm=<no field> split=<no field> maxtop=<no field>. data/table.jsonl is entirely expected_markdown-driven, so one layout annotation on a table document would have erased its headline metric with no error.
2. Extension and extract_field rules crash the load. A rule registered through parse_bench.extensions.register_rule_type validates fine on ParseTestCase and raises ValidationError: 63 validation errors on LayoutDetectionTestCase; an extract_field rule raises too. The exception propagates out of _load_jsonl_dataset, aborting the whole run including unrelated documents.
3. Product-type detection flips for a whole run. _detect_product_type (inference/cli.py) inspects only test_cases[0] of the test-id-sorted list and runs before --group filtering. With a mixed chart/... document sorting first, a llamaparse pipeline's product type is overridden to layout_detection for all five categories.

The class choice buys nothing at evaluation time: _has_mixed_rules accepts both classes and _evaluate_multi_task is typed LayoutDetectionTestCase | ParseTestCase, splitting the rules by rule type, not by test-case class. The layout half is scored either way.

Deliberately not changed

A document whose only parse-side ground truth is an expected_markdown row stays on the layout branch and still loses that markdown — a pre-existing gap. Moving it to ParseTestCase to keep the markdown would be a worse trade: _has_mixed_rules looks for a non-layout rule, finds none, and the document would lose its layout scoring instead. Fixing it properly means teaching _has_mixed_rules about markdown GT, which belongs in its own change.

Two known gaps this does not close, both pre-existing and both only reachable with mixed documents: analysis/comparison.py::_get_gt_annotations accepts only LayoutDetectionTestCase, so a mixed document loses its GT overlay in comparison output; and an extract_field rule on a ParseTestCase is carried but not scored, because ParseEvaluator filters it out.

Scope

The shipped dataset has no document mixing layout with anything else (data/layout.jsonl is 100% type: "layout", and the group key is (category, pdf)), so data/ and data/test/ load identically before and after — verified. No current leaderboard number moves.

Tests

tests/parse_bench/test_cases/test_loader.py keeps @AmirF194's cases and asserts the invariant that matters — no rule of either kind is dropped — rather than the container class. Added: markdown + table settings survive a mixed document, the expected_markdown-only shape stays layout-shaped, and a mixed document with an extract_field rule loads (with an explicit assertion that the same payload is a ValidationError on the other class — the test is about load-time acceptance, not scoring). ruff check, ruff format --check and the test_cases + evaluation + analysis suites (2593 tests) are clean.

[logan-markewich]

this is pretty hard to read lol

[logan-markewich]

Can we shorten to 1-3 lines?

[boyang-zhang1]

Cut to 3 lines each (dc58534), and trimmed the test docstrings the same way. No code change.