feat(k8s): add multimodal Dynamo RL validation

[furionw]

Based on PrimeIntellect-ai/prime-rl#2965.

Why

Prime's native Dynamo backend needs repeatable ARM Kubernetes validation for multimodal rollouts, cross-node trainer handoff, and weight synchronization. Manual runs can reuse stale PVC batches or lose selected GPUs before deployment. This workflow isolates every stage's outputs and deploys the real GPU-requesting pods immediately after just-in-time node selection.

What Change

• Add a digest-pinned, shared-PVC multimodal Kubernetes workflow.
• Add one-step Qwen3-VL-2B and Qwen3.5-2B validation stages plus the longer Qwen3-VL-4B learning stage.
• Use the model-specific qwen3-vl and qwen3.5 renderers so image inputs become vision tokens and training tensors.
• Configure Qwen3.5's required Mamba/NIXL state layout.
• Harden model caching, scheduling, output isolation, and completion checks.

Test Plan

• ARM overlay tests: Dynamo 25 passed; Prime 32 passed, 5 skipped.
• CPU image-render precheck: Qwen3-VL-2B and Qwen3.5-2B both produced image tensors.
• ./run.sh smoke: 16 three-turn image rollouts, 0% errors/truncation; trainer step completed at 256 tokens/s and 18.1 GiB peak memory; orchestrator/trainer exited 0.
• ./run.sh qwen35: 16 three-turn image rollouts, 0% errors and 6.2% truncation; trainer step completed at 222 tokens/s and 20.2 GiB peak memory; orchestrator/trainer exited 0.

Both one-step gates had reward 0 and 0/16 trainable samples; they validate the multimodal RL execution path, not model quality or learning convergence.