fix: align image filenames between parse_html and extract_images when Blank-Page divs present

[voidborne-d]

Problem

parse_html and extract_images generate different filenames for the same image when the document contains Blank-Page divs before Image/Figure divs.

parse_html counts all top-level divs in its div_idx counter (including Blank-Page), while extract_images iterates over chunks from parse_layout/parse_chunks, which filters out Blank-Page divs. This causes the two counters to diverge:

HTML: [Blank-Page(1), Text(2), Blank-Page(3), Image(4)]
                                               div_idx=4 -> hash_4_img.webp

Chunks: [Text, Image]  (Blank-Pages filtered)
               div_idx=2 -> hash_2_img.webp  <-- MISMATCH

The HTML/Markdown output references hash_4_img.webp, but the extracted image is saved as hash_2_img.webp. Broken image link.

This affects any multi-page PDF where the OCR model emits Blank-Page labels for empty or near-empty pages.

Fix

1. LayoutBlock: Add div_idx field to store the original 1-based position among all top-level divs.
2. parse_layout: Use enumerate(top_level_divs, start=1) and store the position in each LayoutBlock.div_idx, preserving the original count even when Blank-Page divs are skipped.
3. extract_images: Read chunk["div_idx"] (carried through parse_chunks -> asdict) instead of maintaining a separate counter.

The change is backward-compatible: div_idx defaults to 0 in the dataclass, and extract_images falls back to idx + 1 if the field is missing.

Tests

17 new tests in tests/test_image_name_consistency.py:

• 5 regression tests -- verify image filenames match between parse_html and extract_images with Blank-Page divs (all 5 fail on the old code, proving the bug)
• 3 baseline tests -- no Blank-Page divs, normal operation still works
• 4 div_idx tracking tests -- verify parse_layout stores correct original positions
• 1 markdown test -- parse_markdown also references the correct filename
• 4 edge cases -- all blanks, missing img tag, empty HTML, no image blocks