[codex] Support raw image refs for multimodal rendering

[eligotts]

Summary

• adds generic mmraw:v2 raw multimodal refs in renderers.mm_store, parsed as RawMMRef objects with family, fingerprint, modality, hash, asset id, and adapter-owned payload
• emits strict prime_raw_mm_item envelopes instead of processed image payloads for Qwen-VL and Kimi K2.5 image rendering
• keeps adapter-specific layout details in renderer-owned payloads (image_grid_thw for Qwen, grid_thws/media token metadata for Kimi)
• supports materializing all raw image refs for retry paths after vLLM multimodal cache misses
• keeps run-scoped image asset refs file-backed so downstream Prime-RL trainer materializes images with its own processor

Companion PRs

• Prime-RL: Support v1 raw multimodal image offload prime-rl#2836
• Verifiers: [codex] Support raw image offload in v1 train client verifiers#1746

Notes

• Draft/WIP: stacked with the Verifiers and Prime-RL raw image offload PRs.
• Verifiers is expected to offload image content to file://.../assets/images/... refs before rendering.
• This intentionally treats raw image refs as the supported path, not processed multimodal feature sidecars.

Validation

• End-to-end hosted-style smoke through Prime-RL with /home/ubuntu/renderers, /home/ubuntu/verifiers, and /home/ubuntu/prime-rl-v1-raw-mm-offload completed inference, env rollouts, train batch creation, trainer step 0, and decoded strict trainer-bound raw image refs.

Note

Support raw image refs for multimodal rendering in Qwen and Qwen3-VL renderers

• Reworks Qwen35Renderer and Qwen3VLRenderer to emit image descriptors (hash, image_grid_thw, placeholder counts) instead of pixel tensors, removing HF processor and per-instance image cache dependencies.
• Adds materialize_image_refs to both renderers and RendererPool, converting descriptor-only image items into run-scoped image references at request time.
• Introduces mm_store.py, a new module providing utilities for run-scoped image reference construction, on-disk asset offloading, and layout fingerprinting.
• Updates generate() in client.py with a materialize_all_image_refs flag; when set, it materializes image refs before request dispatch and builds image-ref selectors for Qwen instead of base64-encoded tensor payloads.
• Replaces image_cache_max with explicit image layout parameters (patch_size, merge_size, min/max_pixels, etc.) in Qwen renderer configs.
• Risk: generate() raises NotImplementedError if materialize_all_image_refs=True and the renderer lacks materialize_image_refs; pixel tensors are no longer embedded in multi_modal_data.

^{Macroscope summarized 32d5a9d. (Automatic summaries will resume when PR exits draft mode or review begins).}