feature: add first-class multi-node multiprocessing deployment support

[observerw]

Describe the feature

Please add first-class support for multi-node multiprocessing deployment in production-stack.

Right now, the distributed Kubernetes path is centered around the RayCluster template:
https://github.com/vllm-project/production-stack/blob/main/helm/templates/ray-cluster.yaml

That template currently hardcodes Ray-based execution, while recent vLLM changes indicate that multi-node deployment should also support the multiprocessing path:
vllm-project/vllm#39400

It would be useful if production-stack exposed a supported way to configure multi-node MP deployment parameters such as nnodes, node-rank, master-addr, master-port, and headless.

Why do you need this feature?

I am trying to deploy a single distributed model across 4 Kubernetes nodes with production-stack.

At the moment, the chart supports the Ray-based path, but does not provide a first-class way to deploy multi-node vLLM with multiprocessing. That makes it difficult to follow the newer vLLM deployment direction without maintaining custom manifests outside of production-stack.

Additional context

Concrete improvements that would help:

• avoid hardcoding Ray-only execution in the current distributed template
• expose multi-node multiprocessing settings through chart values
• provide an official Kubernetes example for multi-node MP deployment

[ruizhang0101]

This feature sounds good! We will take a look at this, meanwhile we welcome any people who would be interested in contributing to this :)