ASR: migrate SenseVoiceSmall to int8 sherpa-onnx runtime (drop PyTorch/funasr) #135

Motivation

Follow-up from #124. The no-GPU ASR default runs SenseVoiceSmall via the funasr
AutoModel path, which pulls in PyTorch + torchaudio and a ~900 MB model.pt.
That stack is heavy and fragile on small/ARM hosts (the #124 crash was a tokenizer
asset that funasr's loader resolved to bpemodel=None), and we have no measured
CPU/ARM real-time factor for the eager funasr path.

The same model served via int8 sherpa-onnx / ONNX Runtime is dramatically
lighter and faster on exactly the hosts that hit #124:

• ~230 MB int8 ONNX weights (vs ~900 MB model.pt)
• no PyTorch dependency — onnxruntime only, with first-class aarch64 wheels
• ~10–20× real-time on ARM Cortex-A76 (sherpa-onnx RK3588 CPU test measured
  SenseVoice int8 at RTF ~0.10→0.05 @ 1→4 threads; the Pi 5 is 4× A76 @2.4 GHz,
  same core family) — i.e. comfortably real-time on a Pi 5 without a GPU
• removes the funasr version-pin + the fun_local.py monkey-patch + the bpemodel
  asset fragility that caused xiaozhi-esp32-server crash looping #124

Scope

• Add a sherpa-onnx-based ASR provider (or adapt fun_local.py) that loads the
  int8 ONNX SenseVoiceSmall export.
• Update make fetch-models to pull the ONNX assets; update make doctor checks.
• Keep language: en behaviour (the existing patch's purpose).
• Benchmark on a real Pi 5 / CPU host and record the RTF before flipping the default.

Notes

• This is the genuine long-term CPU/ARM win identified during the xiaozhi-esp32-server crash looping #124 multi-agent
  analysis. Until it lands, xiaozhi-esp32-server crash looping #124's download fix keeps the funasr/PyTorch path working.
• Decision context: we kept FunASR as the no-GPU default rather than switching to
  WhisperLocal, because Whisper small.en on CPU has the ~5 s/utterance latency
  floor closed in Improve ASR accuracy for children's speech #47. sherpa-onnx SenseVoice is the path that gives both low latency
  and a light dependency footprint.

Refs #124, #47.