Add Fun-ASR-Nano offline ASR

[LauraGPT]

Summary

• add native Fun-ASR-Nano-2512 offline ASR for CPU and CUDA
• add the Kaldi-compatible frontend, SAN-M encoder, adaptor, Qwen decoder, tokenizer, loader, CLI, and server integration
• support official safetensors plus standalone F16/Q8_0 GGUF packages
• publish and pin FunAudioLLM/Fun-ASR-Nano-2512-GGUF@ce72677f84900f0dc57f498ace253bfb3c9155b6
• add model-manager catalog entries, user documentation, and real GGUF transcript regressions

Validation

• CPU Fun-ASR-Nano CTest: 8/8 passed
• CUDA build Fun-ASR-Nano CTest: 15/15 passed
• official checkpoint stage parity: frontend, SAN-M, encoder, adaptor, decoder
• F16 and Q8_0 standalone GGUF transcript parity on CPU and NVIDIA H100 CUDA
• --backend best plus shared Q8 CUDA regression passed
• OpenAI-compatible multipart /v1/audio/transcriptions server E2E returned the exact reference transcript
• loader/catalog sync and git diff --check passed
• independent final review: approved

Runtime Notes

Fun-ASR-Nano is offline ASR only in this PR; streaming and timestamp output are intentionally rejected. CUDA promotes native/F16/Q8 shared decoder storage preferences to BF16 for stable logits while encoder/adaptor weights retain the shared type. An explicit decoder-specific storage option remains authoritative.

Model License

The official checkpoint and converted GGUF weights are governed by the FunASR Model Open Source License Agreement v1.1. The package metadata, model card, and catalog descriptions carry this notice.

[0xShug0]

@LauraGPT Merged! Thanks!

In my test:

backend	path	warm wall ms	RTF	x realtime
CPU	CLI path test	1391.930	0.0989	10.1x
CPU	Server	4585.660	0.3259	3.1x

RTX 5090

backend	path	wall ms	RTF	x realtime
CUDA	CLI path test	158.898	0.0113	88.6x
Vulkan	CLI path test	167.990	0.0119	83.8x
CUDA	Server, threads=8	217.994	0.0155	64.6x
Vulkan	Server, threads=8	234.243	0.0166	60.1x