[MM] Install and load Stable Diffusion 3 models

[lstein]

Summary

This implements alpha-level support for generating images with Stable Diffusion 3.

To use:

1. Install the Diffusers version of SD3, located at stabilityai/stable-diffusion-3-medium-diffusers. For convenience, this repo has been added to the Starter Models pane of the model manager UI.
2. Create a workflow like the one shown below:
   
3. If you are on a CUDA system with less than 20 GB of VRAM, deactivate the "Use TE5 Encoder" slider in the Stable Diffusion 3 node. Otherwise you will run out of memory!

Related Issues / Discussions

Currently only diffusers models can be loaded. Upstream support for loading checkpoints is a WIP.

• Probe SD3 diffusers
• Probe SD3 checkpoints
• Load SD3 diffusers
• Load SD3 checkpoints
• Load SD3 diffusers models with a quantized text_encoder_3
• Unload SD3 diffusers models with a quantized text_encoder_3 (not confirmed working)
• Download and install SD3 diffusers
• Download and install SD3 checkpoints

Models using a Quantized text_encoder_3

There is preliminary support for such models, but they still do not fit into a 12 GB VRAM machine and this feature has not been tested. Systems with adequate VRAM may be able to load and run these models, but be aware that there is an unsolved issue with an 8K VRAM memory leak that occurs with each generation.

Note that a quantized model will only run on CUDA systems.

QA Instructions

Follow the instructions above.

Merge Plan

Squash when approved.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• Documentation added / updated (if applicable)