fix: prevent LoRA weight accumulation in Flux loader

[Cor-net]

Summary

Fixes LoRA weight accumulation in complex workflows (e.g. multiple NunchakuFluxLoraLoader nodes or dynamic LoRA switching in UI).

This version:

• ✅ Removes unsafe copy.deepcopy(model) — fixes the __setstate__ crash (TypeError: 'NoneType' object is not callable, AttributeError: 'pinned') that occurs in ComfyUI v0.4.0+ when deepcopying a LoRA-modified ModelPatcher.
• ✅ Introduces nunchaku_base_model — a clean clone stored on first use — to ensure all LoRA applications start from the same base.
• ✅ Explicitly resets transformer LoRA state (reset_lora(), comfy_lora_meta_list = [], comfy_lora_sd_list = []).
• ✅ Removes duplicate LoRA entries by path before applying new ones.

Testing

Fully verified in real-world scenarios:

• ✅ Single NunchakuFluxLoraLoader (e.g. Turbo LoRA),
• ✅ Two chained loaders (Turbo → Style), with strength changes and LoRA swaps,
• ✅ NunchakuFluxLoraStack,
• ✅ Dynamic LoRA switching in UI → no cached generations, no noise after 5+ changes.

🔍 Note on workflow behavior: In ComfyUI’s DAG execution, changing only the second LoRA node does not re-execute the first (cached). Full isolation is guaranteed when using a single LoraStack or when the entire chain is re-triggered (e.g. via new seed). The node-level fix ensures correctness regardless of workflow topology.

Fixes #716
Tested on: ComfyUI v0.4.0, ComfyUI frontend v1.34.8, Windows 11, CUDA 13.0.