add support for LoRA weights

[ssube]

https://huggingface.co/docs/diffusers/training/lora#inference

[ssube]

This is working for some weights, like the reference at https://huggingface.co/sayakpaul/sd-model-finetuned-lora-t4, but failing for many of the files tagged as LoRA on Civitai with an error about missing metadata: cloneofsimo/lora#191

[ssube]

I have this working for both weights produced by https://github.com/cloneofsimo/lora and https://github.com/kohya-ss/sd-scripts, like https://civitai.com/models/8039/jackscape-samurai-jack-background-style-lora:

This is not a one-step process yet and relies on both of those repositories, so there's quite a bit of work to still be done, but it is possible.

[ssube]

This is now documented: https://github.com/ssube/onnx-web/blob/main/docs/converting-models.md

I'm not sure if the one-step conversion will make v0.8 or come later, but it can be done manually and the resulting ONNX models can be shared.