feat(inference): normalize native routed experts

[biswapanda]

Summary

• normalize native terminal routed-expert NumPy payloads into Prime-RL's existing replay envelope at the renderer boundary
• validate base64 size, NumPy headers, exact byte counts, rank-three shape, integer dtype, byte order, and expert-index bounds before allocation or replay
• preserve the existing legacy dictionary envelope unchanged
• install the wrapper idempotently in environment workers so trainer replay receives the normalized route data

This is the routed-expert increment following #3481, #3482, and #3483. Until those dependencies merge, this fork PR is cumulative against main; it remains draft until the exact routed-experts Kubernetes gate passes.

Validation

• Ruff and Python compilation passed on all changed files
• focused Dynamo adapter and native routed-payload tests: 86 passed
• independent correctness and security review of bounded NPY parsing and worker-hook behavior
• the frozen routed source passed local tests and produced Kubernetes binaries; the native preflight and exact five-step routed-expert training gate are pending and are not claimed as passed

Dependencies

• feat(trainer): support VLM training with dynamo  #3483
• feat(vllm): forward routed experts data ai-dynamo/dynamo#13531
• feat(rust): return routed experts through gRPC vllm-project/vllm#52723