SDXL + SDXL Turbo UNetLoader, LoRA, FB Cache, ControlNet support

[Pizzawookiee]

Adds the nodes 'NunchakuSDXLUNetLoader', 'NunchakuSDXLLoraLoader' and 'NunchakuSDXLLoraStack'. Does not modify the base nunchaku library.

The file models/sdxl.py converts the state dictionary from the existing nunchaku SDXL checkpoint on huggingface into one that is compatible with ComfyUI's UNetModel, which is distinct from the diffusers 2DUNetConditionModel.

Compare this image generated with the new node (should be able to reproduce my workflow by dragging it into ComfyUI):

With one copied from https://comfyanonymous.github.io/ComfyUI_examples/sdturbo/:

Update 25/10/29:
LoRA nodes implemented. Includes "NunchakuSDXLLoraLoader" and "NunchakuSDXLLoraStack". Because I converted everything to work with ComfyUI's UNetModel (and not diffusers 2DUNetConditionModel used in the base nunchaku repo), my LoRA implementation does not require any update to nunchaku library.

LoRA and LyCORIS should work. DoRA is not supported right now, I'm not sure how to apply the magnitude_vector DoRA layer to the quantized qweight Nunchaku layer.

From my testing, most LoRAs work but some don't. For example both the Lightning and LCM LoRAs cause horrible image quality.

Additionally, it appears ControlNet sort of works (can use standard ComfyUI Controlnet nodes, and sometimes get okay results), but see Known Issues at bottom.

Update 25/10/30:
First-block cache implemented.

Known issues:

1. In most ControlNet workflows I run into a "CUDA error: an illegal memory access was encountered".