fix censor behavior for japanese diactritics

[dgisser]

Restore Japanese Kana With Dakuten

• Added a regression test showing that Japanese kana like パピプペポ / ぱぴぷぺぽ currently get normalized to their unvoiced forms (ハヒフヘホ / はひふへほ) by the diacritic stripping pass.
• Updated the normalization filter (src/censor.rs:159-175) so the combining dakuten (\u{3099}) and handakuten (\u{309A}) marks are preserved while still removing other nonspacing marks and banned characters. This keeps Japanese kana intact without loosening other filtering behavior.

Testing

Before fix:

cargo +nightly test japanese_diacritics_preserved -- --nocapture
...
thread 'censor::tests::japanese_diacritics_preserved' panicked
  left: "パピプペポ"
  right: "ハヒフヘホ"

After fix:

cargo +nightly test japanese_diacritics_preserved -- --nocapture
running 1 test
test censor::tests::japanese_diacritics_preserved ... ok

Full suite:

• cargo +nightly test censor

[finnbear]

Looks great, thanks! I did initially suggest detecting Japanese (like Devanagari) but this is a better solution as there are only 2 diacritics and they're hard to abuse. This change will take effect in our games next time I update the filter.

[dgisser]

for sure, let me know when it pushes to prod and I'll test it out!