feat(vllm-omni): preserve generated video audio

[furionw]

Based on #13706.

Why

Joint audio-video diffusion models return video frames as the primary output and audio plus media rates as multimodal metadata. Dynamo currently reads only the frames and re-encodes a silent VP9 MP4, losing a model-generated modality. The formatter should preserve every generated video, mux synchronized audio when present, and leave the existing royalty-free silent-video path unchanged.

What Change

• Normalize batched tensor and array video layouts without dropping outputs.
• Mux generated audio into H.264/AAC MP4 responses with rate metadata.
• Add an opt-in codec overlay with a real encode-and-reopen smoke test.

classDiagram
    DiffusionFormatter ..> MediaMuxer : muxes joint outputs
    DiffusionFormatter ..> VideoData : records actual media metadata
    DiffusionFormatter --> MediaStorage : uploads completed MP4 bytes

    note for DiffusionFormatter "Normalizes, pairs, encodes, and formats diffusion outputs"

Loading

Test Plan

• Run formatter tests for silent, joint, tensor, and batched outputs.
• Build the codec overlay and validate both MP4 streams.