[examples/controlnet/train_controlnet_sd3.py] training SD3.5 large controlnets from stability

It is currently not possible to fine-tune controlnets for SD3.5 large from stability ai like stabilityai/stable-diffusion-3.5-large-controlnet-blur due to the handling of positional embeddings.

I have extended and testet the script accordingly since I need it for my use case. Would this be interesting and should I do a PR?

@sayakpaul

Start with examples/controlnet/train_controlnet_sd3.py and compare its positional-embedding handling with stabilityai/stable-diffusion-3.5-large-controlnet-blur. Reproduce the SD3.5 Large fine-tuning path, then verify that the script supports the cited controlnet model without positional-embedding errors.

Vom Indexierungsmodell aus dem Issue-Text verfasst.