fix: limit vLLM VAD segments to 30 seconds

[LauraGPT]

Summary

• configure the standalone Fun-ASR-Nano vLLM server's FSMN-VAD model with a 30-second maximum segment duration
• add a regression test that verifies the limit is applied when the VAD model is constructed

Why

The server constructed FSMN-VAD without max_single_segment_time, so it inherited the 60-second default. Passing this value to AutoModel.generate() is ineffective because FSMN-VAD stores it in VADXOptions at model construction time.

On the 480-second attachment from #3419, a 60-second segment omitted a substantial passage around the Blackwell/product update and robot discussion. The 30-second configuration matches the existing Fun-ASR-Nano examples and documentation.

Validation

• python3 -m pytest tests/test_fun_asr_nano_autocast_device.py tests/test_fun_asr_nano_ctc_batch_fallback.py tests/test_fun_asr_nano_missing_ctc_weights.py tests/test_fun_asr_nano_openai_response.py tests/test_fun_asr_nano_repetition_penalty.py -q (27 passed)
• python3 -m py_compile examples/industrial_data_pretraining/fun_asr_nano/serve_vllm.py tests/test_fun_asr_nano_openai_response.py
• python3 -m black --check tests/test_fun_asr_nano_openai_response.py
• git diff --check origin/main...HEAD

Real-audio comparison using the issue attachment:

Metric	Before	After
VAD segments	13	20
Maximum segment duration	60.01 s	30.01 s
Aligned, script-normalized CER	22.74%	19.63%
RTF	0.0087	0.0169

The CER comparison removes the unmatched 0-17.69 second teaser that is absent from the reference transcript and converts both reference and hypothesis to Simplified Chinese before retaining letters and numbers. The recovered passage is present after the change, although proper-name recognition such as Blackwell remains imperfect.

Related to #3419.

[gemini-code-assist]

Caution

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