Add experimental support for externalizing graphs to a separate artifact

[jakesabathia2]

• Adds two experimental TorchConverter SPI hooks that act on the ops a named nn.Module class
  contributed, identified by nn_module_stack attribution once the graph body is built:

API	Effect
_set_delegate_id(class_name, target)	put all ops in the subgraph into a single coreai.isolated_group<target> graph
_set_externalize_group(class_name, namespace, *, externalize=True)	moves them into their own coreai.graph (marked externalize, inside namespace) and replaces them with a coreai.invoke
_set_externalize_name_map(mapping)	pins outlined graph symbols when the traced module path is not the name the artifact uses

• Important to note that the SPIs are experimental and no guaranteed of not breaking in the future.
• I did try to integrate the feature with the existing externalize APIs - however, the solution become super tedious and need tons of model reauthorizing at the source torch model level - hence I decided to introduce this new approach 😄 Need more discussion to put this into production.
• Make debug info infra faster