feat(workflow): transfer remote tensors with NIXL

[furionw]

Based on #13293.

Why

Most workflow values should use Dynamo's ordinary endpoint codec. A remote stage that produces a tensor needs a different transport and ownership lifecycle. This optional fork enables NIXL at binding time, selects it from the runtime value, and keeps the carrier opaque to graph authors.

Experimental: This workflow API is under active development and may evolve based on feedback. It is an optional orchestration layer over Dynamo's discoverable workers and endpoints; applications may compose those primitives directly when they prefer to own the orchestration logic.

What Change

• Add tensor_carrier="nixl" as an opt-in remote binding capability.
• Select NIXL only for a top-level runtime torch.Tensor; ordinary values keep the endpoint codec.
• Derive consumers from WorkflowIR and bindings instead of adding physical edges to the plan.
• Retain producer storage until graph-derived consumers complete and quarantine uncertain transfers.
• Reject unsupported workflow-input, inline-boundary, workflow-output, and nested tensor cases.

classDiagram
    RemoteBinding --> NixlTensorCarrier : enables capability
    RemoteStageServer --> NixlTensorCarrier : exports runtime tensor
    NixlTensorCarrier *-- NixlLeaseRegistry : retains producer
    NixlTensorCarrier ..> NixlTensorFanout : creates consumers
    RemoteStageClient --> NixlCarriedValue : imports opaque carrier message

    note for RemoteBinding "WorkflowIR remains transport-neutral"
    note for NixlLeaseRegistry "Releases only after planned consumers finish"
    note for NixlCarriedValue "Internal; stage author receives tensor"

Loading

Test Plan

• GPU-free tests cover value selection, ordinary fallback, fan-out, lifecycle completion, timeout quarantine, invalid topology, and direct mappings.
• Applicable pre-commit hooks pass.