feat: add FunASR dataset annotation tool

[LauraGPT]

Is this PR adding new feature or fix a BUG?

Add feature.

Is this pull request related to any issue? If yes, please link the issue.

Closes #1291.

Summary

• add an optional FunASR dataset annotation CLI that recursively creates Fish Speech .lab files from WAV, MP3, and FLAC audio
• preserve curated labels by default, with explicit dry-run and overwrite modes, atomic writes, single-file failure isolation, and deterministic statistics
• keep FunASR outside the core dependency surface through lazy imports and document CUDA/CPU usage in English and Chinese

The default SenseVoiceSmall path writes plain, one-line UTF-8 transcripts for Mandarin, Cantonese, English, Japanese, and Korean. Model selection, device, language, and ITN remain configurable.

Validation

• python3 -m unittest discover -s tests -v (9 passed)
• python3 -m compileall -q fish_speech tools tests
• all repository pre-commit hooks passed on the four changed files
• dependency-free --help and --dry-run smoke passed without importing FunASR
• real H100 smoke with FunASR 1.3.27 and the official SenseVoiceSmall Chinese and English samples: 2 processed, 0 failed; exact expected transcripts, one line each, with control tags removed
• second-run skip and --overwrite paths both verified; clean validation environment reports no broken requirements
• git diff --check and credential scan passed

No Fish Speech runtime dependency or training behavior changes are included.

[LauraGPT]

@leng-yue, could you review this optional dataset-annotation tool when convenient? You recently merged changes in the same installation/finetuning documentation area.

The current head is CLEAN and mergeable, with no unresolved review threads or failed/pending checks. FunASR stays lazily imported outside Fish Speech's core dependencies; the CLI preserves existing .lab files by default and has dry-run, overwrite, atomic-write, and per-file failure coverage. Exact-head validation includes 9 unit tests, repository pre-commit hooks, dependency-free help/dry-run smokes, and real H100 SenseVoiceSmall transcription for the official Chinese and English samples.