fix(optimizer): preserve Unicode heading keys

[rudycelekli]

Flash expansion currently strips every non-ASCII letter from its comparison keys. Printed Japanese subsection titles then become empty strings and are rejected as identical to the parent; an absent non-ASCII title can also match any page beneath an ASCII parent. Preserve Unicode letters and numbers while retaining the existing lowercase, whitespace, punctuation and underscore handling.

Add four public page_index_flash regressions using generated Japanese and ASCII PDFs. They exercise native PDF parsing, bookmarks and optimization, with only external completion responses stubbed. The cases check printed heading extraction, duplicate/parent/absent-title rejection, and page ranges when a subsection starts at the top or midway through a page.

Validation:

• Four focused native PDF cases pass; restoring the old normalization expression makes all four fail.
• Full local suite: 589 passed, 218 skipped on Python 3.12/PDFium 5.13 without optional agent frameworks. This used the same production source; the focused cases and mutation check were repeated after moving the printed-text assertion earlier in the tests.
• git diff --check passed. No real model calls were made.