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

[furionw]

Based on #3.

Why

The single-node Dynamo launcher owns every process on one host, but Prime's multi-node Slurm template previously could not coordinate one frontend, role-specific workers, shared discovery, and service teardown across ranks. Multi-node aggregate and P/D RL need the parent allocation to own those processes without nested Slurm submissions, while standalone Dynamo inference remains unsupported. Shared-file discovery and explicit completion markers keep worker identity stable and prevent long-lived services from holding a finished allocation.

What Change

• Launch one frontend and role-indexed Dynamo worker per inference node.
• Validate shared-filesystem and supported aggregate/P/D topology shapes.
• Coordinate clean cross-node completion and document runnable examples.

Test Plan

• Single-node gate: aggregate and P/D passed three steps with zero effective errors.
• Run multi-node aggregate and P/D E2Es before marking ready.

Gate evidence: 20260717T181440Z/portable-evidence/summary.json passed both topologies with zero fatal signatures.