feat(pdfrx_engine): restore logical order to Hebrew text stored in visual order

[palmoni5]

Many scanned Hebrew books come with an OCR text layer that stores each line's words left to right. Some older fonts go further and store every word's letters left to right too. Either way, loadText/loadStructuredText return the text backwards, so phrase search never matches and copied text comes out reversed.

This adds PdfHebrewTextNormalizer, opt-in through Pdfrx.normalizeHebrewText (default false).

How it decides. It uses the character geometry, not guesswork:

• A word whose Hebrew letters advance left to right is reversed. Combining points stay after their letter.
• A line whose Hebrew words advance left to right is reordered right to left. Whitespace keeps its slots.
• A word or line with too little evidence (a single pair) follows the page majority, so short lines on a logical page are never flipped.
• Text already in logical order is returned untouched (the same instance).

Legacy encodings. It also decodes Windows-1255 Hebrew that a font without a ToUnicode map exposes as Latin-1, or as Mac Roman. This only happens on a page where those characters outnumber the Latin letters, so French or German text is never touched.

Character boxes are permuted together with the text, so indices into the result keep addressing the right glyphs; highlighting and selection keep working. The PDFium backend runs the normalizer on the worker.

Measured on real books

• 15 random HebrewBooks scans with a text layer: 7 had reversed word order, with 0–13% of adjacent word pairs right-to-left. After the fix they measure 87–100%.
• The 8 scans that were already correct are left byte-for-byte unchanged.
• In a 39-tractate Talmud set, only the two broken files change. One of them has its Windows-1255 / Mac Roman runs decoded, about 76k letters over 8 pages. The other 37 are untouched.

Unit tests cover logical text, word-order reversal, full visual order, points, Latin-1 decoding, accented Latin text, and the page-majority rule.

[espresso3389]

@palmoni5 Very impressive and interesting work. I'll look into it.