Support v-prediction checkpoint models when set in the model manager

[dunkeroni]

Summary

Capture prediction type from model settings and pass it to the scheduler to enable v_prediction checkpoints to work again.

Related Issues / Discussions

I'm certain there is a more correct way to handle this in the model manager that would also allow automatic detection from checkpoint files, but this at least gets vpred models working again before we go make sweeping changes to the model probe code.

Since our conversion code does not check the settings on the checkpoint, attempting to convert a v_prediction model to Diffusers will still default over to epsilon prediction and it will forever fail afterwards. It is not currently possible for users to change a diffusers model to v_prediction in Invoke without editing the local json files. User Mewt on discord has found that huggingface repos with correct scheduler_config.json are able to work without issues, but I have found others online that still specify epsilon.

Closes #7495
Closes #6268 (edit by @psychedelicious to get GH to pick up that this PR closes this issue)

QA Instructions

Before:

After:

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[GrettaKlu]

Not sure what the right way of doing this is, but the changes here error when using the upscale tab. The file "tiled_multi_diffusion_denoise_latents.py" on line 216 references the get_scheduler() function and needs the unet_config argument passed through as well.

[RyanJDick]

I added a fix for the Multi-Diffusion inference issue in 160a36a.

I tested the following...

Setup:

1. Install noobaiXLNAIXL_vPred10Version model
2. The prediction_type gets probed incorrectly as epsilon (this will have to be addressed in a separate PR)
3. Change the model's prediction_type in the models tab.

Tests:

• Generate with the noobai model (v-pred), and verify that it produces coherent results
• Generate with an epsilon-pred model and verify that the output is the same as before this change.
• Upscale with the noobai model and verify that it works in a Multi-Diffusion pipeline
• Upscale with an epsilon-pred model and verify that the output is the same as before the change.

[psychedelicious]

I believe this also closes #6268.