feat(vllm-omni): add MiniMax-H3 request adapter - #13708

Based on #13707.

Why

MiniMax-H3 is available in Dynamo's pinned vLLM-Omni dependency, but Dynamo does not express its startup or request contract. H3 requires model-aware 24-FPS and shape defaults, task inference across two checkpoint partitions, H3-specific sampling controls, and strict reference validation. These behaviors belong in a model adapter so generic video generation remains unchanged.

What Change

• Forward H3 startup topology, attention, text-encoder, and offload controls.
• Map task-aware sampling defaults while preserving upstream task inference.
• Enforce matching H3 request validation in Rust and Python.

classDiagram
    OmniConfig --> BaseOmniHandler : supplies H3 startup controls
    BaseOmniHandler --> AsyncOmni : constructs combined H3 pipeline
    OmniHandler --> OmniDiffusionSamplingParams : maps request controls
    OmniHandler --> OmniTextPrompt : attaches H3 reference lists

    note for OmniHandler "Applies model-aware H3 defaults and task contract"

Test Plan

• Run H3 protocol, argument, handler, and reference-contract tests.
• Exercise task-less T2VA, FL2VA, and Ref2VA input shapes.