fix(text-chunking): split Default-mode CJK text at full-width punctuation

[gqf2008]

Summary

split_word_ranges in the Default text-chunking mode cut words only at ASCII spaces. CJK text has no inter-word spaces, so an entire paragraph parsed as a single word; once the text exceeded the codepoint budget (text_chunk_size, default 200) the hard-split fallback kept the whole paragraph as one chunk and the budget was never applied.

Every TTS family that goes through chunk_text_request (omnivoice, outetts, heartmula, moss_voicegen, audio8_tts, ...) is affected, in both offline and streaming runs: long Chinese text synthesized as one oversized chunk regardless of text_chunk_size. text_chunk_mode=tag_aware was the only workaround.

Fix: full-width CJK sentence/clause punctuation (。！？，、；：) now forms word boundaries — the punctuation token becomes a word of its own carrying its sentence/clause break flags, so the existing greedy budget + rollback logic splits at natural boundaries. ASCII punctuation stays attached to its word, leaving Latin text byte-identical.

Verification

New text_chunking_test (5 cases: CJK over-budget splits at sentence boundary, under-budget stays whole, clause-punctuation rollback, Latin sentence split unchanged, oversized single-word passthrough). Confirmed on macOS arm64 and across the full CI matrix (Linux cpu/vulkan, macOS, Windows):

• 236-codepoint CJK paragraph with the default 200 budget: audio8_tts.text_chunk_count 1 → 2
• 115-codepoint short text: output wav byte-identical to the pre-fix build
• Latin text: chunks identical to pre-fix

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release -DENGINE_BUILD_TESTS=ON
cmake --build build --target text_chunking_test
./build/text_chunking_test   # text_chunking_test: all passed

[0xShug0]

@gqf2008 Merged. Thanks!