Fix nondeterministic PDF markdown with batch_size > 1 (#2319)

[Kylinny]

Fixes #2319.

Hi, and thank you @rtorrero-lw for the excellent report — the diagnosis and local repro made the root cause easy to confirm.

Bug

With the default batch_size=4, NaivePDFProcessorStrategy.process_batch() produced different markdown for the same PDF on different runs. In the issue's repro (12 pages, each with a Title Case line that clean_pdf_text() only bolds on page 1), 30 runs gave 5 distinct outputs locally (bolding landed on page 1, on no page, or on other pages); batch_size=1 was always stable.

Root cause

Each worker thread wrote its page number to the shared self.current_page_number attribute and _process_page() read it back after text extraction — by which time another worker could have overwritten it. The same stale value also fed PDFPage.page_number and the extracted image filenames (page_{n}_img_{m}, which could even collide across threads).

Fix

Pass the page number down explicitly instead of storing it on self, exactly as suggested in the report: _process_page(page, image_dir, page_num + 1) and _extract_images(page, image_dir, page_number). The now-unused current_page_number attribute is removed. Invariants preserved:

• 1-based page numbering is unchanged on both the serial process() and batched process_batch() paths (verified: both produce identical per-page markdown/html).
• clean_pdf_text() / clean_pdf_text_to_html() signatures are untouched; the page-1-only formatting now provably applies only to page 1.
• Remaining self reads inside worker threads (extract_images, save_images_locally, …) are all read-only after __init__.

Tests

New tests/unit/test_pdf_page_number_race.py (hand-crafted multi-page PDFs, no new test dependencies):

• test_process_batch_is_deterministic — 10 runs with batch_size=4 give byte-identical output
• test_page_one_only_formatting_in_batch — author-line bolding appears on page 1 and only page 1
• test_page_numbers_sequential_in_batch — PDFPage.page_number matches page position under threads
• test_process_page_uses_given_page_number (parametrized) — _process_page honors its argument
• test_process_and_process_batch_agree — serial/batch parity

All 7 new cases fail on the unfixed code and pass with the fix; the existing test_pdf_html_escaping.py PDF tests still pass.

Before / after (issue's repro, 30 runs, batch_size=4)

• Before: 5 distinct outputs over 30 runs (bolding on pages [1], [], [1, 2], [1, 4], [1, 6])
• After: 1 distinct output over 30 runs — bolding on page 1 only, every time

Happy to adjust anything on review. Thanks again for the great catch!