[SDXL-IP2P] Update README_sdxl, Replace the link for wandb log with the correct run

[priprapre]

What does this PR do?

It replaces the wandb run with the correct link in README_sdxl.md for instructpix2pix training. The new link for wandb run is from the diffusers/sdxl-instructpix2pix-768 model page. (The old one was pointing to SD1.5 instructpix2pix run)

I think there is a typo on the model page as well. It says that batch size per GPU is 8 (x4 gradient accumulation=32) but in the wandb run, I see that it's 4 (x4 gradient accumulation=16).

Besides, I observe that some parameters differ from the script on README_sdxl. For example, the diffusers/sdxl-instructpix2pix-768 model is trained with learning_rate=5e-06 and conditioning_dropout_prob=0.1 whereas the script defaults are 5e-05 and 0.05. Hyperparameters differ from SD1.5 IP2P training, so it might be useful to change those defaults for SDXL so that we would know which set of hyperparameters worked out well for SDXL case.
Maybe @harutatsuakiyama also has some insights here about the hyperparams working the best.

cc: @sayakpaul

[sayakpaul]

Thanks!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.