fix: align SenseVoice sentence timestamps

[LauraGPT]

Summary

• align SenseVoice VAD text, punctuation tokens, and word timestamps without treating unspaced Chinese as one token
• preserve exact surface text for contractions, URLs, and email addresses while merging BPE timestamp units
• keep CT punctuation text and punc_array consistent when forcing a final sentence terminator

Fixes #3406

Verification

• pytest -q tests/test_punc_model_none.py tests/test_generate_subtitle.py (12 passed)
• focused tests/test_auto_model.py regression selection (4 passed, 1 deselected)
• py_compile, Black check, and git diff --check
• real 45.1-second SenseVoice Chinese audio: 12 subtitles, 0 empty, 0 punctuation-only, max 21 characters, final timestamp 44.680s, no length-mismatch warning

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.