RAE+DiT support for published checkpoints

**Is your feature request related to a problem? Please describe.**
`diffusers` has `AutoencoderRAE` support via #13046. There is still no native latent generator path for the published RAE checkpoints. I've been following the research and was excited to see the RAE support; this comes next and would be exciting to have!

Already, the official RAE repo ships a custom Stage-2 DiT-like latent generator and public checkpoints/configs, but today these are not available through a native `diffusers` model + pipeline + checkpoint conversion flow.

**Describe the solution you'd like.**
In two stages, probably: (a) inference support for the checkpoints (b) training support.
I have some code ready that I've been able to use to get the checkpoints workable, could turn it around into a PR.

Start with the AutoencoderRAE support from #13046 and inspect the official RAE repo's Stage-2 DiT-like latent generator, including its published checkpoints and configs. Define the native diffusers model, pipeline, and checkpoint conversion flow for inference first, then assess the separate training-support stage. Done means the published checkpoints work through a native diffusers path.

Written by the indexing model from the issue text.