In-chain overlays

[martin-traverse]

This allows flows to be overlaid without creating a new flow definition or splitting the flow into separate jobs.

Overlays are attached to flow sockets as a list, multiple overlays can be applied to each socket and are executed in order.

Overlays can be applied to any socket, including flow inputs and outputs as well as input and output sockets for models. Overlays applied to a flow input or model output affect all consumers of that dataset. Flows applied to a model input or flow output only affect a single dataset.

Implemented by running a graph transform that expands the original graph by inserting the overlays, then executing the expanded graph in the regular executor. Implementation effort is surprisingly small! Overlay nodes are tagged to provide more meaningful logging and error reporting.

Originally done at the platform level in the orchestrator, but implementing in the runtime would make overlays available during model build and via the TRAC shell.