Autoscaler-driven elastic expert/data parallelism for MoE

Feature Description and Motivation

For large MoE models we only scale whole replicas, which is coarse and expensive. vLLM has elastic EP (enable_elastic_ep, /is_scaling_elastic_ep, the DP coordinator request-wave protocol in vllm/v1/engine/coordinator.py) to add/remove DP/EP ranks at runtime. The autoscaler doesn't drive it. Scaling EP width by load would right-size MoE serving without full-replica steps.

Use Case

DeepSeek / Qwen-MoE style deployments where per-replica cost is high and traffic varies; scale expert parallelism up and down instead of adding or removing entire replicas.

Proposed Solution

Let PodAutoscaler target EP/DP width for elastic-EP-enabled deployments and drive scaling via the engine's elastic-EP endpoints + the DP coordinator wave protocol. Start observe-only (read /is_scaling_elastic_ep and wave state) before acting on it.

Start by reading the elastic-EP references in vllm/v1/engine/coordinator.py and review the /is_scaling_elastic_ep endpoint and DP coordinator request-wave protocol named in the issue. Trace how PodAutoscaler currently targets replicas, then define how observe-only wave-state reporting and later EP/DP width changes would be validated for elastic-EP-enabled deployments.