Improve table detection and header refinement in find_tables()

[veget-able]

This follows #5057 and addresses cases where find_tables() misses tables, detects chart grids as tables, or merges separate tables. It also improves header tagging for spanning cells and reduces repeated work during extraction.

• Virtual lines and boxes. Forward add_lines and add_boxes to the nested line finder when union=True, so they participate in table detection.
• Batched grid lines. In lines_strict, retain thin filled rectangles longer than edge_min_length even when many are grouped into one fill path. The path's overall bounding box is excluded from the grid. This is the largest individual gain on ParseBench, concentrated in one document family that draws tables this way.
• Tables inside picture regions. Filter union candidates at the cell-group stage, before building a Table. Candidates mostly inside a layout picture region must have text distributed across at least two rows and two columns, each at least half populated, with mostly rectangular cell coverage. This helps reject chart grids while retaining tables misclassified as pictures.
• Stacked tables. With refine=True, split a grid when a confirmed leading header row repeats verbatim. The header must match the column structure, and each resulting table must retain at least one row below the header.
• Headers below spanning cells. With refine=True, extend the header when the next row supplies individual column labels beneath a cell spanning several, but not all, columns and does not match the body pattern. At least one body row remains. This rule changes header_rows and <th>/<td> tagging in to_html() without changing bounding boxes or extract() output.
• Extraction cost. Compute Table.header on first access using the original detection call's extraction flags. Reuse vector drawings within each find_tables() call and index word centres per page to speed up word selection.

On the 1,463-page regression corpus, output is byte-for-byte unchanged in both the default and lines_strict configurations. With refine=True, one page changes because its column-label row is now included in the header. This PR adds 16 tests in tests/test_tables.py.

On the ParseBench table group (503 pages, pymupdf4llm_markdown pipeline), this PR together with the companion pymupdf4llm reading-order change raises GTRM from 72.00 on PyMuPDF 1.28.2 to 75.64. The header-completion rule contributes 0.58 points, with 10 pages improving and none regressing. For the other changes, 34 pages improve and two score lower. Those two cases are a chart grid that previously received partial credit and a datasheet header block inside an image, both now rejected by the picture-region filter.

Output on the held-out DP-Bench set (200 pages, 55 tables, evaluated with TEDS) is unchanged from 1.28.2. With both PRs applied, total wall time on the 503-page ParseBench set is 13% lower than 1.28.2. Processing the additional thin rectangles can add up to 0.85 seconds on individual pages containing thousands of them.

[veget-able]

The two Ubuntu jobs fail in tests/test_widget_reproducers.py (the widget reproducer tests added to main on 23 September), not in the table tests, which all pass. The same tests fail on a plain 1.28.2 install here as well, so they look independent of this PR. Happy to rebase again once those are addressed on main.