feat(dynamo): integrate sidecar inference

[biswapanda]

Summary

This is the integration branch for end-to-end review and validation of the split Prime-RL changes needed by the Dynamo vLLM sidecar path. It combines:

• vLLM 0.26 token-serving compatibility
• checkpoint-boundary preservation
• external-engine rank mapping
• Dynamo worker discovery
• native NCCL/NIXL weight transfer
• filesystem LoRA lifecycle
• Dynamo example recipes

Split PRs

• refactor(vllm): delegate token serving to vLLM 0.26 #3174 vLLM 0.26 token serving
• fix(nccl): preserve checkpoint layer boundaries #3175 checkpoint layer boundaries
• fix(inference): map external engine ranks #3177 external-engine ranks
• feat(inference): Dynamo integration with prime-rl #3176 Dynamo discovery
• feat(dynamo): support native weight transfer #3178 native weight transfer
• feat(dynamo): support LoRA lifecycle #3179 Dynamo LoRA lifecycle
• docs(dynamo): add training recipes #3180 Dynamo examples

The split PRs are the intended review and merge path. This draft provides a single branch for integration CI and end-to-end testing and should not be merged independently of that review sequence.

Compatibility

The Dynamo path remains opt-in through dynamo_discovery_url. Existing Prime-RL static, elastic, and Python vLLM paths remain the default. The branch is restacked on current upstream main, including the latest Verifiers and research-environments updates.

Test plan

• Ruff passed on all changed Python files
• 37 focused serving, discovery, topology, NCCL/NIXL, LoRA, orchestrator, and config tests passed
• broader configuration collection reached optional tasksets absent from the reused local environment; Dynamo-specific configuration tests passed