fix(subtitles): split overlong source cues

[LauraGPT]

Summary

• split a single overlong model sentence into bounded readable subtitle cues
• preserve model word surfaces and per-token timestamps so English and multilingual cues retain real word/audio boundaries
• pass aligned sentence words through both the CLI and subtitle example
• keep the original model segment when timestamps are missing, malformed, unordered, unsupported, or contain an indivisible over-limit token instead of fabricating timing

Verification

• PYTHONPATH=$PWD python -m pytest tests/test_cli.py tests/test_generate_subtitle.py -q -> 27 passed
• python -m compileall -q funasr/cli.py examples/subtitle/generate_subtitle.py tests/test_cli.py
• exact 10m33s reporter sample on H100: 84 cues with 26 over 8s before; 131 cues after, maximum 7.979s / 42 characters, with 0 over-limit cues, 0 overlaps, and 0 leading/trailing whitespace
• repeated fresh inference produced SHA256 7617b91b3ef8ce23b650a38a6a52b7c03f1e8f0f15b82ce5d8caabd55db286d1
• independent review: no remaining P1/P2 findings
• git diff --check

Addresses the overlong-cue boundary reported in #3539. The issue remains open until the reporter verifies the result on the affected mixed-language recording.