feat(workflow): add integrated encoder ensemble

[furionw]

Based on #13293.

Why

The integrated-encoder workflow keeps multimodal encoding inside stock vLLM while composing an independent classifier. Its application-owned response join normally belongs beside the orchestrator, but deployments may isolate that final composition through discovery. Both placements should use one authored graph and the same response-stage implementation.

What Change

• Fan the opaque request value to stock vLLM and a remote classifier.
• Bind response composition inline or through a discovered endpoint.
• Launch the response worker only for remote placement.

classDiagram
    WorkflowOrchestrator --> StageDispatcher : executes plan
    StageDispatcher ..> GenerateEndpointInvoker : calls stock vLLM
    StageDispatcher ..> RemoteStageClient : calls remote stages
    RemoteStageClient --> DummyClassifier : classifies request
    StageDispatcher --> EnsembleResponseStage : joins inline or remotely

    note for WorkflowOrchestrator "Coordinates generator, classifier, and join"
    note for EnsembleResponseStage "Shapes the frontend token chunk"
    note for RemoteStageClient "Discovers classifier and optional response worker"

Loading

Test Plan

• 49 cumulative workflow tests pass; two environment-specific tests skip locally.
• CPU coverage includes graph fan-out, exact name wiring, and both response placements.
• pre-commit passes for the PR diff.

[github-actions]

🌿 Fern Docs Preview: https://nvidia-preview-01a026f3-04d2-723c-868b-a98c5c08f153.docs.buildwithfern.com/dynamo/dev