Explicit graph building function for adding sub-graph

[EwanC]

The record & replay graph building API lets a user add a sub-graph as a node to the graph by recording a queue queue::submit(command_graph<executable>) operation. However, there is no equivalent mechanism in the explicit graph
building API for adding a sub-graph.

CUDA defines cudaGraphAddChildGraphNode in it's explicit API which does similar.

We should consider adding a function like command_graph<modifiable>::add(command_graph<executable/modifiable>, const std::vector<node> &dep = {}) to achieve this functionality.

[EwanC]

It was decided from discussing #26 which changes graph submission to handler::exec_graph(command_graph<executable>) that this design would also allow us to add an executable graph as a sub-graph. The implementation of executable graph can contain all the necessary information from the original graph it was created from to do this correctly and enable optimizations.

However, we may still want to have a way modifiable graph added as a sub-graph, which would match cudaGraphAddChildGraphNode. The downside of this however is that the runtime would have to clone the modifiable graph when adding it as a sub-graph node, to avoid any future user additions being propagated to the parent graph.

Leaving this issue as open as we should analyze application usage for both explicit and record&replay mode, to see created a sub-graph from a modifiable graph is a gap we need to fill