[RFC sub-task] Migrate magi_human example to standard task example + model_extras

[princepride]

Sub-task of #4539.

Migrate the magi_human model off its dedicated special-usage example onto the standard task example + a declarative model_extras registration, following the reference migration in #4225 (BagelPipeline / SenseNovaU1Pipeline).

Scope

• Special-usage example dir(s): examples/offline_inference/magi_human/

Steps

1. Add vllm_omni/model_extras/magi_human.py declaring EXTRA_BODY_PARAMS / EXTRA_OUTPUT_PARAMS, and register it in vllm_omni/model_extras/registry.py.
2. Route the pipeline's previously-hardcoded knobs through extra_body / custom_output (keep a legacy fallback during the transition).
3. Delete the bespoke example dir(s) above; fold usage into the relevant standard task example (text_to_image / image_to_image / text_to_video / text_to_audio / etc.) with an extra_body recipe for any non-default params.
4. Update e2e test(s) to exercise magi_human through the standard path and pass in CI.
5. Update markdown/README: remove the per-model README, document the knobs in the task README/recipe.

⚠️ Mandatory for the PR (will be sent back if missing)

• A complete, copy-pasteable, reproducible run script (model id + all extra_body params + env vars)
• The actual run results/output pasted into the PR (generated image/video/audio + logs; request+response for serving)
• Updated e2e test passing in CI
• Updated markdown/README

Definition of done

• vllm_omni/model_extras/magi_human.py added + registered
• Pipeline knobs routed via extra_body / custom_output
• Bespoke example dir deleted; usage folded into standard task example
• e2e test updated + passing
• README/markdown updated
• Reproducible script + actual results in PR description