Fix Chatterbox multilingual diacritic normalization

[0xShug0]

Summary

Fixes #462.

Chatterbox multilingual text preprocessing now uses the framework NFKD decomposition helper instead of the llama tokenizer NFD helper. The previous helper mapped each input codepoint to a single output codepoint, so precomposed diacritics such as Polish accented letters could lose their combining marks before tokenization.

This keeps the fix scoped to Chatterbox multilingual tokenization and does not change the English tokenizer path.

Validation

• Built audiocpp_cli.
• Reproduced the issue with Polish NFC input before the fix.
• Verified that, after the fix, Polish NFC input and the same pre-decomposed input produce byte-identical audio with the same seed.
• Verified the fixed NFC output with Qwen3 ASR; the transcription preserves Polish diacritics.