Serverless executor

Turned out to need no new mechanism — just checking that task-scoped execution genuinely isn’t tied to one platform or one coordination model.

duckpipe.run(module, only=task, state_uri=...) is the reference executor — a plain Python call with no platform-specific glue. ../examples/07_serverless_executor proves it by dispatching one DAG’s two tasks into the same distributed run through two genuinely different invocation shapes: a container (docker run, argv-invoked — ECS/Cloud Run/a Kubernetes Job call the same way) and a handler(event, context) function (the FaaS calling convention Lambda/Modal/Cloud Functions actually use). Neither pipeline.py nor DuckPipe itself knows or cares which one dispatched it.

For one big task instead of many independent ones, see remote_execution.md — “beefy node” mode needs none of this machinery at all.