[Frontend] Add content_parts to /inference/v1/generate for raw multimodal input

[aoshen02]

Summary

• Add content_parts field to GenerateRequest for raw multimodal input on the /inference/v1/generate endpoint
• Both Rust and Python frontends resolve media server-side — no pixel data transfer needed
• Enables RL frameworks to pass token_ids + raw media (URLs/base64) in a single request

Motivation: RL workloads

RL frameworks (e.g. prime-rl) need token-level inference with multimodal inputs. Today's options:

Path	Problem
/v1/completions	No multimodal support in OpenAI spec
/v1/chat/completions	Requires messages + chat template — RL callers already have token_ids
/render → /generate	Pixel data round-trip (~MB per image, see RFC vllm-project#46722)
/generate + content_parts	Token IDs + raw media → one request, zero overhead

Usage

POST /inference/v1/generate
{
  "token_ids": [1, 32010, 32044, ...],
  "content_parts": [
    {"type": "image_url", "image_url": {"url": "https://example.com/photo.jpg"}}
  ],
  "sampling_params": {"max_tokens": 64, "temperature": 0}
}

Design

• Reuses OpenAI content part format (image_url, input_audio, video_url, etc.) — same schema as chat completions, consistent across Rust and Python frontends
• Rust path: content_parts → content_parts_to_media_parts() → ChatLlm::prepare_media() → MmFeatures on TextRequest
• Python path: content_parts → AsyncMultiModalItemTracker → TokensPrompt → OnlineRenderer pipeline
• Mutually exclusive with features (pre-processed path from /render)
• 6 files changed, ~125 lines added

Test Plan

• Unit test: Rust content_parts_to_media_parts conversion
• E2E test: Python TITO endpoint with content_parts (image URL → generation)
• E2E test: Rust frontend with content_parts

Resolves vllm-project#51472

🤖 Generated with Claude Code