fix: support Fun-ASR MLT Nano inference

[LauraGPT]

Summary

• recognize FunAudioLLM/Fun-ASR-MLT-Nano-2512 as a Fun-ASR Nano model
• reuse the Nano inference profile (cache, batch_size, and list-based hotwords)
• avoid loading the separate ct-punc model for MLT Nano

Why

The current model-name check only matches the contiguous fun-asr-nano string. The
official multilingual model ID contains fun-asr-mlt-nano, so it falls back to the
classic FunASR profile.

That profile adds ct-punc, passes batch_size_s, and uses the singular hotword
argument. Fun-ASR Nano reads the list-based hotwords argument instead, so configured
hotwords are silently lost for MLT Nano.

The official MLT model configuration declares model: FunASRNano, and the official
inference example uses cache={}, batch_size=1, and hotwords=[...]:

• https://huggingface.co/FunAudioLLM/Fun-ASR-MLT-Nano-2512/blob/main/config.yaml
• https://github.com/modelscope/FunASR/tree/main/examples/industrial_data_pretraining/fun_asr_nano

Verification

• python -m unittest discover -s tests -p "test_*.py" -v (15 passed)
• ruff check . with the repository-pinned ruff==0.15.21
• python -m compileall -q app tests
• git diff --check

[LauraGPT]

Marked ready for review after refreshing validation on head 2ddd329724bfd306a434d0352582839724a16f50.

Validation:

• python -m unittest tests/test_core.py -q -> 15 tests passed
• python -m compileall -q app main.py tests/test_core.py
• git diff --check

The added coverage verifies that both FunAudioLLM/Fun-ASR-Nano-2512 and FunAudioLLM/Fun-ASR-MLT-Nano-2512 use the Nano inference options, while traditional FunASR/SenseVoice models keep the legacy hotword string path.

[LauraGPT]

Validated the current head 2ddd329724bfd306a434d0352582839724a16f50 against current upstream main@5da86bcb.

Checks run:

• python3 -m unittest tests.test_core -v -> 15/15 passing
• python3 -m py_compile app/services/asr_funasr.py tests/test_core.py main.py
• git diff --check

The new test coverage includes the MLT checkpoint path (FunAudioLLM/Fun-ASR-MLT-Nano-2512) and verifies it uses the Nano-style inference options (batch_size=1, per-call cache, list-form hotwords, no batch_size_s / classic hotword). GitHub currently reports the PR as mergeable; the UNSTABLE label is from the absence of status checks, not a failing check.

[harry0703]

Thank you for contributing Fun-ASR MLT Nano support and for adding focused regression coverage. I have merged this after verifying the change locally with the real MLT checkpoint and an end-to-end audio transcription. The model now initializes and transcribes successfully through AudioNotes. I will handle a couple of follow-up refinements separately, including deterministic decoding and clearer handling for timestamp-based output, so this contribution can remain focused. Thanks again!