Add CPU offloading support for all diffusion models

[LawJarp-A]

Purpose

Add CPU offloading support for all diffusion models in vllm-omni. This enables memory-efficient inference by automatically moving model components (text encoder, DIT transformer, VAE) between CPU and GPU based on usage, reducing GPU memory requirements.

Key features:

• Alternating offload strategy: offload DIT during text encoding, offload text encoder during DIT inference
• Configurable via existing config flags (dit_cpu_offload, text_encoder_cpu_offload, vae_cpu_offload)
• Enabled by default for all diffusion models

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft.

BEFORE SUBMITTING, PLEASE READ https://github.com/vllm-project/vllm-omni/blob/main/CONTRIBUTING.md

[hsliuustc0106]

is there any similar features in diffusers? if any, please link related issues or PRs

[ZJY0516]

is there any similar features in diffusers? if any, please link related issues or PRs

https://huggingface.co/docs/diffusers/main/en/api/pipelines/overview#diffusers.DiffusionPipeline.enable_sequential_cpu_offload

[ZJY0516]

@LawJarp-A will write a RFC soon