feat(video): support typed input references

[furionw]

Based on #13705.

Why

The video API exposes only one untyped image reference, while vLLM-Omni pipelines can condition generation on image, video, and audio inputs. Dynamo needs a typed contract that preserves per-modality order, materializes remote media within request-level limits, and adapts singleton values to existing pipeline ABIs. Backends without this capability must reject the field instead of silently producing unconditioned video.

What Change

• Add typed image, video, and audio references to Rust and Python protocols.
• Materialize validated references with count, byte, and cleanup boundaries.
• Scalarize generic singleton inputs and reject unsupported backend usage.

classDiagram
    NvCreateVideoRequest --> VideoReferenceMaterializer : supplies typed references
    VideoReferenceMaterializer --> OmniHandler : returns pipeline-ready paths
    OmniHandler --> OmniTextPrompt : attaches multimodal data

    note for VideoReferenceMaterializer "Bounds, stages, groups, and cleans request media"
    note for OmniHandler "Adapts references to the selected pipeline ABI"

Loading

Test Plan

• Run Rust/Python video protocol and materializer tests.
• Run SGLang and TensorRT-LLM rejection tests.