feat(vllm): complete Prime RL protocol integration

[biswapanda]

Summary

• expose model-scoped RL worker discovery and sidecar-mediated control routes for separately deployed Prime-RL trainers
• support native NCCL, NIXL/ModelExpress, and filesystem compatibility updates without introducing a Prime-specific receiver protocol
• preserve policy-version cache salts and carry typed, preprocessed multimodal features through exact request routing and vLLM-RS lowering
• add generic routed-experts request and terminal response transport through Dynamo and the vLLM sidecar
• synchronize the vendored inference protocol with the exact vLLM-RS contract used for validation

Base and relationship to earlier work

This PR is intentionally based on acc5ee96e39c4c84506f87aa2fd15589682deb4e, the exact head of ai-dynamo#13066 used for end-to-end validation. It is a fresh review delta in this fork and does not update or reuse #3.

The multimodal routing design selectively reuses the generic capability and routing contracts from ai-dynamo#11588 and ai-dynamo#11532 while retaining the newer native sidecar implementation from ai-dynamo#13066.

Impact

Prime-RL can discover and control a stable Dynamo worker cohort, generate through aggregated or prefill/decode deployments, update independently deployed vLLM-RS workers, and train with multimodal or routed-experts data while the ordinary Dynamo paths remain unchanged unless the new fields and routes are used.

Validation

• cargo fmt --all -- --check
• cargo test -p dynamo-vllm-sidecar (33 unit tests plus the executable doctest)
• cargo test -p dynamo-llm protocols::openai::generate (28 tests)
• cargo test -p dynamo-kv-router mm_identifier_hash_preserves_vllm_and_opaque_identifiers
• native token-generation, RL discovery/control, filesystem reload, and post-update generation smokes
• three-step NIXL/ModelExpress training
• three-step aggregated native-NCCL training on separate trainer and inference GPUs
• three-step 1P/1D native-NCCL training with directly observed KV handoff
• three-step exact Qwen/Qwen3-VL-4B-Instruct multimodal training and post-update generation
• three-step routed-experts capture/replay training and post-update generation