docs(dynamo): add training recipes

[biswapanda]

Summary

• document the minimal Prime-RL configuration for a Dynamo sidecar deployment
• add Qwen3 0.6B Math, Qwen3 30B Thinking, and GLM-5.2 FP8 R2E recipes
• keep deployment-specific Kubernetes wiring outside the Prime-RL Helm chart

Dependency

• feat(inference): Dynamo integration with prime-rl #3176 introduces the dynamo_discovery_url consumed by these recipes

This branch includes the prerequisite commit so the examples remain configuration-valid while targeting main.

Test plan

• ruff check on the combined changed surface
• Dynamo-specific configuration resolution tests
• focused aggregate validation passed; full optional-recipe parametrization requires the repository's taskset extras

---

Note

Medium Risk
Changes weight-broadcast topology validation and NCCL initialization paths for external inference; misconfigured inference_world_size or discovery could deadlock or skip broadcasts, but scope is additive behind dynamo_discovery_url.

Overview
Adds external Dynamo inference support so Prime can train against a separately managed Dynamo/vLLM stack without launching [inference].

Runtime & config: Orchestrator clients can set dynamo_discovery_url (mutually exclusive with admin_base_url and elastic DNS). A new DynamoInferencePool polls /v1/rl/workers, validates worker snapshots, and wires per-engine admin URLs for weight updates. Shared weight_broadcast.inference_world_size is required for Dynamo orchestrator configs and is propagated through RL resolution so NCCL/NIXL (and filesystem for LoRA) know how many inference ranks to expect; single-node NCCL no longer demands two local GPUs when that external world size is set explicitly.

Docs/examples: examples/dynamo/ documents the K8s/env overlay (RL ports, sidecar POD_IP, NCCL worker extension) and adds smoke recipes for Qwen3-0.6B math, Qwen3-30B Thinking (1P/1D), and multi-node GLM-5.2 FP8 R2E with split trainer/orchestrator TOMLs.

^{Reviewed by Cursor Bugbot for commit 2f67c72. Bugbot is set up for automated code reviews on this repo. Configure here.}