fix: load official SenseVoice ONNX models

[LauraGPT]

Fixes #101.

Summary

• detect model_quant.onnx and pass quantize=True to funasr_onnx.SenseVoiceSmall
• keep custom non-quantized model.onnx directories working with quantize=False
• validate the required config, CMVN, tokenizer, and ONNX files before importing/loading the runtime
• document the exact two-repository ModelScope preparation flow in Chinese and English

The official iic/SenseVoiceSmall-onnx repository ships model_quant.onnx but does not include chn_jpn_yue_eng_ko_spectok.bpe.model, while the existing loader used the constructor default quantize=False. This made the documented SenseVoice path fail even after downloading the official ONNX repository.

Validation

• 59 passed with the locked project environment (uv sync --extra web)
• focused quantized, non-quantized, and missing-tokenizer regression tests
• Ruff check and format check on changed Python files
• Python compile check and git diff --check
• verified the actual funasr-onnx==0.4.1 wheel constructor supports quantize: bool = False
• verified both referenced ModelScope repositories are reachable

Validation Refresh (2026-09-29)

Rechecked unchanged head 27006d703898c7447369b745fb052f66b5618c44 against main 0248c84593bf66391dc2e38b4583ff72e3eae456 in the project's locked Linux/Python 3.12.3 environment:

• uv sync --frozen --extra web --extra sensevoice completed successfully, including the actual optional runtime dependencies. No lockfile changes.
• Actual funasr_onnx.SenseVoiceSmall and postprocessor imports succeed with funasr-onnx==0.4.1, torch==2.11.0, and onnxruntime==1.24.4.
• uv pip check --python .venv/bin/python: all 95 installed packages compatible.
• uv run --frozen --no-sync --extra web --extra sensevoice pytest -q -ra: 59 passed, 6 dependency warnings, no skipped tests.
• git diff --check: passed; repository source and PR head unchanged.
• Both the minimum supported 0.4.0 wheel and locked 0.4.1 wheel support the quantize constructor argument. Downloaded library wheel digests were checked against PyPI metadata.
• Fresh official ModelScope file listings confirm the documented split: the ONNX repository contains model_quant.onnx and supporting config/CMVN/token files, while the tokenizer .bpe.model must come from iic/SenseVoiceSmall. Config and CMVN metadata hashes match across the two repositories.

These are environment, import, unit-test, runtime-contract, and remote metadata checks. No model weights or tokenizer artifacts were downloaded, no model was loaded, and no audio inference or GPU validation was performed. Earlier Ruff/format/compile checks above are historical, not rerun as part of this refresh.

Separate Pre-existing ITN Gap

Issue #105 tracks a separate issue already present on main: the adapter sends use_itn, but locked funasr-onnx==0.4.1 reads textnorm instead. A no-weights probe against the actual runtime selects woitn / 15 for both adapter settings, while an explicit textnorm="withitn" control selects 14. Existing passing tests do not cover that transcription argument contract. This PR does not fix ITN selection; its scope remains model-directory preparation and quantization.