feat(slurm): support single-node Dynamo RL jobs

[furionw]

Based on #2964.

Why

PrimeIntellect-ai#2964 adds a Dynamo inference backend, but a top-level [slurm] currently propagates into the nested inference config, where Dynamo is rejected. Single-node RL does not need a separate inference submission: after Slurm assigns the node, the batch job can enter Prime's existing local Dynamo launcher. This enables scheduled aggregate and P/D workflows while preserving the rejection for standalone inference and multi-node Dynamo deployments.

What Change

• Keep single-node Dynamo local inside Prime-managed Slurm allocations.
• Preserve Dynamo worker topology in generated inference configs.
• Add aggregate and P/D examples, documentation, and regression coverage.

Test Plan

• Run focused config and RL entrypoint unit tests.
• Complete three-step aggregate and P/D workflows on one four-GPU Slurm node.

Validated with ai-dynamo/dynamo#11254.