feat(protocols)!: support cached multimodal UUID content parts

[furionw]

Why

vLLM extends OpenAI-compatible chat content parts with opaque multimodal cache UUIDs, including UUID-only requests where the media field is null. Dynamo needs to represent that wire contract without removing its existing nested UUID API. The canonical top-level field is explicitly documented as vLLM-only; legacy nested UUID fields remain available but deprecated.

ref: https://docs.vllm.ai/en/stable/features/multimodal_inputs/#image-inputs_1

Effect

{"type":"image_url","image_url":null,"uuid":"catalog-image-42"}

What Change

• Add top-level opaque UUIDs and nullable media content.
• Preserve required nested URLs and deprecated UUID fields.
• Cover URL-backed, UUID-only, and legacy nested payloads.

Test Plan

• cargo test -p dynamo-protocols
• cargo check --workspace --all-targets

[keivenchang]

pre-approving, but please address the change-of-type comments

[devin-ai-integration]

Devin Review found 0 new potential issues.

[rmccorm4]

Can't find any particular reason to keep uuid nested within the current fields.

I don't see it defined in https://github.com/64bit/async-openai nor https://developers.openai.com/api/reference/resources/chat/subresources/completions/methods/create#(resource)%20chat.completions%20%3E%20(model)%20chat_completion_content_part_image%20%3E%20(schema)

I see SMG defined it in a similar level as the existing location, nested in image_url dict: https://github.com/lightseekorg/smg/blob/d43aa0420fe7e30504801e3af77d26d0f59826f7/crates/multimodal/src/types.rs#L52, but I don't see it used in their documentation or examples either, it is always set to None: https://github.com/lightseekorg/smg/blob/d43aa0420fe7e30504801e3af77d26d0f59826f7/model_gateway/src/routers/grpc/multimodal/detect.rs#L38

Similarly I don't see dynamo using the uuid field in the current location with any value other than None: https://github.com/search?q=repo%3Aai-dynamo%2Fdynamo+%22uuid%22+image_url+language%3ARust&type=code

Let's go with this to support the vllm uuid feature. Thanks for adding the deprecation notice.

[rmccorm4]

Further context here: https://docs.vllm.ai/en/stable/features/multimodal_inputs/#image-inputs_1