feat: add Qwen encoder runtimes (Qwen2.5-VL + Qwen3) for Nunchaku

[tonera]

Summary

This PR adds Nunchaku runtimes for Qwen-family text encoders and exposes them via the public nunchaku API. It includes:

• A stable entry point (NunchakuQwenEncoderModel) that auto-dispatches to the correct runtime based on exported .safetensors metadata.
• Dedicated runtimes for:
  • Qwen2.5-VL text-only encoder (NunchakuQwen2VLTextEncoderModel)
  • Qwen2.5-VL multimodal edit encoder (NunchakuQwen2VLEditEncoderModel)
  • Qwen3 text-only encoder (NunchakuQwen3TextEncoderModel, optional depending on transformers version)

Key changes

• Public exports
  • Export Qwen encoder models from nunchaku/__init__.py and nunchaku/models/__init__.py
  • Add nunchaku/models/text_encoders/__init__.py exports
• New Qwen encoder runtimes
  • nunchaku/models/text_encoders/qwen_encoder.py: dispatcher from_pretrained() selecting the correct runtime via checkpoint metadata (text_encoder_usage, model_type, runtime_class)
  • nunchaku/models/text_encoders/qwen2_vl_text_encoder.py: Qwen2.5-VL text-only runtime (rejects multimodal args)
  • nunchaku/models/text_encoders/qwen2_vl_edit_encoder.py: Qwen2.5-VL edit/multimodal runtime
  • nunchaku/models/text_encoders/qwen3_text_encoder.py: Qwen3 text-only runtime
  • nunchaku/models/text_encoders/qwen_common.py: unified base class, checkpoint metadata parsing, config building, quantized linear patching, and rotary buffer materializers
• Transfer utilities
  • nunchaku/torch_transfer_utils.py: device normalization + pin-memory policy + optional “pretouch” helpers for CPU→CUDA staging

Files changed

• Updated (overwritten)
  • nunchaku/__init__.py
  • nunchaku/models/__init__.py
  • nunchaku/models/text_encoders/__init__.py
• Added
  • nunchaku/models/text_encoders/qwen_encoder.py
  • nunchaku/models/text_encoders/qwen_common.py
  • nunchaku/models/text_encoders/qwen2_vl_text_encoder.py
  • nunchaku/models/text_encoders/qwen2_vl_edit_encoder.py
  • nunchaku/models/text_encoders/qwen3_text_encoder.py
  • nunchaku/torch_transfer_utils.py

Test plan

• Import smoke test:
  • python -c "from nunchaku import NunchakuQwenEncoderModel, NunchakuQwen2VLTextEncoderModel, NunchakuQwen2VLEditEncoderModel; print('ok')"
• Dispatcher path:
  • python -c "from nunchaku import NunchakuQwenEncoderModel; m=NunchakuQwenEncoderModel.from_pretrained('path/to/qwen_encoder.safetensors', device='cpu'); print(type(m))"
• (Optional) Qwen3 availability:
  • If transformers provides Qwen3Model, load a Qwen3 exported checkpoint; otherwise confirm it raises a clear ImportError message.

Model download link:

https://huggingface.co/tonera/Qwen2.5vl-Nunchaku
https://huggingface.co/tonera/Qwen3-text-Nunchaku