Add llama.cpp / GGUF runtime for Fun-ASR-Nano (CPU / edge)

[LauraGPT]

Fun-ASR-Nano on llama.cpp / GGUF

Run Fun-ASR-Nano (SenseVoice SAN-M encoder + adaptor + Qwen3-0.6B) entirely on the
llama.cpp / ggml stack — CPU / edge, a single binary, no Python at runtime. Like
whisper.cpp, but for Fun-ASR.

What's added (runtime/llama.cpp/)

• funasr-cli — integrated WAV → transcription binary
• funasr-encoder / funasr-embd — the ggml encoder+adaptor and the LLM-from-embeds tools (validation)
• export_encoder_gguf.py — export the audio encoder + adaptor to GGUF (f32 / f16)
• README with architecture, quickstart, accuracy and gotchas

How it works

fbank (C++) → SAN-M encoder + adaptor (ggml) → low-frame-rate truncation →
[prefix tokens | audio embeds | suffix tokens] → Qwen3-0.6B (llama.cpp). The audio
embeddings are injected via llama_decode's embedding-input path (the llava/mtmd mechanism).

Validation (184-file benchmark)

• Encoder + adaptor (ggml) vs PyTorch: cosine 1.000000, max_abs_diff 5e-3 (f32)
• kaldi fbank (C++) vs torchaudio: cosine 1.000000
• End-to-end CER, identical conditions (f32 LLM, 15s chunking): C++ micro 11.68% vs PyTorch 11.70% (Δ0.02%)
• Best practical (Q8 LLM + 15s chunk): micro-CER 9.51%

Notes

• Use --chunk 15 for long audio. WAV input assumes 16 kHz mono PCM16 for now.
• Does not modify any existing code; everything is under runtime/llama.cpp/.