[RFC]: Support multi-node serving in CI

[pkking]

Motivation.

Currently, all CI workflows run on a single Ascend server node, which limits the maximum available NPU count to 8 cards.

But its impossible to test multi-node scenarios witch maybe more close to real world use case.

This RFC aims to provide a solution that enables community developers to write test cases for multi-node vllm serving deployments.

Proposed Change.

Since the community CI runs on a kubernetes cluster, there are many out-of-box multi-node serving solution, for example lws and in vllm project there's also an example for reference, the straightforward idea is to build by lws directly, here's a general plan:

1. add a new workflow, which contains two jobs:
   1. job1: create a new lws instance which expose a vllm service with multi pods on seperatly node
   2. job2: wait the lws service is ready, then run the tests, job must cleanup the resource when tests finished
2. add some guides to help developer how to setup more multi-node style test cases

since multi-node serving may be time and NPU comsuming, It's best not to triggered by a PR

Feedback Period.

Maybe one week

CC List.

@Yikun @wangxiyuan @Potabk @MengqingCao

Any Other Things.

No response

[Yikun]

Feel free to open the PR, I'd like to review once it's ready.

[Potabk]

Maybe we need to list the scenarios where we most urgently need in the CI.

• From the perspective of deployment(deployment difficulty ranges from easy to complex):
  1. multi-node on ray backend
  2. multi-node on mp backend
  3. disaggregated pd(maybe cover the second point)
• From the perspective of models
  1. deepseek series
  2. qwen3-moe series
• How frequently we need to test them and how should developers to trigger them or just a scheduler job for now