Train from inline multimodal rollouts

[eligotts]

Summary

• accept raw multimodal content_parts on the pinned vLLM token endpoint and return its effective expanded prompt_token_ids
• build on main's renderer-free orchestrator path and trace-stamped mm_token_type_id_map, carrying only inline image URLs and expanded placeholder ranges through transports.batch
• preserve raw image refs through truncation, packing, and workload splitting so distributed runs avoid unnecessary dummy multimodal batches
• materialize one trainer microbatch at a time through a small model-family adapter registry with explicit forward and context-parallel policies
• support Qwen-VL and Kimi K2.5 while leaving the existing multimodal SFT path unchanged
• pin Renderers PR length rewards #133 and Verifiers PR fix(ring_attn): make FA3 causal/window_size kwargs version-agnostic #2417 at 36f8636f3

Stack

• Renderers: Add raw multimodal content parts support renderers#133
• Verifiers: Keep multimodal content inline in v1 rollouts verifiers#2417
• 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/unit/inference/test_serving_tokens.py tests/unit/orchestrator/test_batch.py tests/unit/orchestrator/test_qwen3_vl_e2e.py tests/unit/train/test_multimodal.py tests/unit/train/test_model_forward.py (44 passed)
• Ruff check and format passed