fix(auto-model): align SentencePiece word markers

[LauraGPT]

Summary

• ignore the SentencePiece ▁ word-boundary marker while aligning ASR word timestamps with punctuation tokens
• add a regression test covering punctuation-preserving sentence timestamps for marked words

Root cause

The reporter's mixed-language sample contains one timestamped word prefixed with ▁. The visible ASR surface correctly omits that tokenizer marker, but _merge_timestamp_units compared it as transcript text. That single-character mismatch invalidated alignment for the whole recording, triggered the VAD-segment fallback, and discarded punctuation from sentence_info even though the top-level text contained 272 punctuation marks.

Validation

• python -m pytest -q tests/test_punc_model_none.py tests/test_cli.py tests/test_generate_subtitle.py -> 45 passed
• reporter's 633.651 s sample on exact head 83de4060de043f9ea0a0c318edc9a818c54f28d5:
  • 3,173 words and timestamps; the sole ▁ marker now aligns with the punctuation surface
  • sentence_info: 275 entries carrying all 272 predicted punctuation marks
  • readable SRT: 137 cues, maximum 7.94 s and 42 characters
• 224.427 s comparison sample:
  • sentence_info: 153 entries carrying all 152 predicted punctuation marks
  • readable SRT: 53 cues, maximum 6.36 s and 42 characters

Related to #3539. This PR intentionally does not auto-close the issue; the reporter should confirm the fix on their workflow first.