Add raw multimodal content parts support

[eligotts]

Summary

• add an opt-in raw multimodal path that preserves renderer-owned text and tool formatting without running image processors
• send ordered raw media as content_parts to the vLLM token-generation endpoint
• use the effective prompt_token_ids returned by vLLM while retaining the logical renderer prompt for strict RL extension
• preserve the existing eager multimodal path by default, including SFT callers

Stack

• Verifiers: Keep multimodal content inline in v1 rollouts verifiers#2417
• Prime RL: Train from inline multimodal rollouts prime-rl#3320
• Upstream vLLM content-parts support: [Frontend] Add content_parts to /inference/v1/generate for raw multim… vllm-project/vllm#51478

Validation

• uv run pytest -q tests/test_client.py (25 passed)

Note

Add raw multimodal content parts support to client.generate and multimodal renderers

• Adds raw_multimodal flag to client.generate; when true, a new _content_parts helper flattens image_url, audio_url, and video_url parts from messages into content_parts and sends them in the POST body
• When raw multimodal is used, the engine response must include prompt_token_ids; the return now carries renderer_prompt_ids (renderer-side tokens) and prompt_ids (engine-provided tokens). A missing prompt_token_ids raises MalformedGenerateResponseError
• Marks KimiK25Renderer, Qwen35Renderer, and Qwen3VLRenderer with supports_raw_multimodal = True and adds a process_multimodal flag to their render and bridge_to_next_turn methods; when false, placeholder tokens are still emitted but multimodal tensors and metadata are skipped
• Risk: renderers called with process_multimodal=False no longer populate mm_hashes, mm_placeholders, or mm_items; any caller relying on those structures for tensor construction must handle the absent data

^{Macroscope summarized 26f4ea3.}