Support Qwen3 ASR 1.7B HF checkpoints

[mirek190]

Summary

• add support for loading official Qwen/Qwen3-ASR-1.7B-hf checkpoints
• support the flat Hugging Face configuration, processor metadata, tokenizer format, tensor namespaces, and tied language-model output used by the 1.7B model
• retain compatibility with Qwen3-ASR 0.6B
• update the model manager and documentation for the new model

Why

The official Qwen3-ASR 1.7B Hugging Face checkpoint differs from the existing 0.6B layout and could not be loaded directly by audio.cpp.

Validation

• built and tested the Windows CPU CLI
• built and tested the Windows CUDA CLI and server
• inspected both Qwen3-ASR 1.7B and 0.6B checkpoint layouts
• ran real Japanese WAV inference on an RTX 3090
• tested the managed server and model-manager workflows

[0xShug0]

@mirek190 Thanks for the PR! I’ll test it, and if everything looks good, I’ll merge it.

[mirek190]

No problem

I test it and seems working fine.
Tested with English and Japanese.
Also tested with Qwen3-ForcedAligner as a timestamp.

1.7b is much better than 0.6b

[0xShug0]

@mirek190 merged! Thanks again for your contribution!