Support 1-n mapping of model_key to deployment in NDIF (model multireplica).

[khaiwang]

Enable model replicas (one-to-many mapping)

Goal

• Support multiple concurrent replicas of the same model. First stage is strictly the 1 model_key to N replicas mapping and a naive request routing method; smart scheduling and auto‑scaling are out of scope.

Scope

• Add a replica identity (replica_id) and use it in actor naming.
• Allow controller/cluster to track and create multiple deployments per model_key.
• Allow the processor to route requests across multiple replicas.
• Update status/CLI to surface replica information.

Functional requirements

• A model can be deployed with a specified replica count (e.g. replicas=3).
• Each replica is a distinct Ray actor named ModelActor:{model_key}:{replica_id}.
• Dispatcher/Processor can send requests to any available replica of the model.
• The request will be routed based on simple statistics of ongoing requests.
• Status APIs and CLI show replicas per model_key and their state.
• Backward compatible: if replica_id isn’t provided, default to 0.

Non‑goals (stage 1)

• No autoscaling, no placement optimization, no replica health‑based scheduling beyond basic
  selection.

[khaiwang]

#237