Add x-prediction / prediction_type=sample support for SDXL fine-tuning

[bghira]

What does this PR do?

Adds x-prediction / prediction_type=sample support.

I sort of worked it out how to do this on my own, would love feedback.

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline?
• Did you read our philosophy doc (important for complex PRs)?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[patrickvonplaten]

Ok for me - @sayakpaul wdyt?

[sayakpaul]

Super!

Any exciting results you wanted also to share benefitting from this objective?

[bghira]

i havent gotten the chance to train an x-pred model ... yet =) maybe someone will beat me to it now

[bghira]

The reason I added sample prediction training support is due to the results in the Bytedance paper, "Common Diffusion Noise Schedules and Sample Steps are Flawed" that demonstrated similar visual acuity to velocity objective. The paper noted that x-prediction results tend to be blurrier than v-prediction, but that both are of higher quality than noise (epsilon) prediction.