feat: improve furigana filter

[Kellenok]

1. Hybrid Furigana Filter: Replaced the old OwOCR logic with a global filter that checks precise geometry and orientation.
2. Massive Accuracy Improvement: Achieved 98.8% accuracy (F1-score) on the benchmark dataset vs 89% before, reducing total errors by nearly 9x (false negatives dropped from 77 down to 4, and false positives from 19 down to 7).
3. Faster Processing: Implemented regex-based "early exits" to skip Kanji and alphanumeric strings, making the filter 20% faster (1.19ms vs 1.47ms per page) despite the more thorough checks.

[Kellenok]

merger became slightly worse, i'll try improve it tomorrow

[sohilsayed]

Try to also resolve conflicts

[Kellenok]

there was more slop code than i expected, sorry

from 100 random manga109 images
before this pr: 64 overmerged, 59 undermerged
after furigana: 60 overmerged, 68 undermerged
after merging fix: 25 overmerged, 25 undermerged

but i checked kuso onna and some other manga and can't say that the results were exactly twice as good as on manga109, just better in some places, worse in others. i don't get it and i'm tired

[Kellenok]

before / after

also need to test local engine, i can't now

[sohilsayed]

honestly looks good enough for me we can always keep it for future work