Improve %load_node for node that doesn't have persisted dataset

[noklam]

Description

Originated from kedro-org/vscode-kedro#140

%load_node works great but it doesn't work on MemoryDataset, I don't save every node and it's not easy to figure out which nodes I need to run again to produce the data

Context

• This is much more powerful because not every node has persisted data and this limited the usage of the feature.
• This enable much more powerful slicing feature in kedro-viz, which is currently limited because we do not want the slicing generate a command that Kedro does not know how to run.
• Same as above, if we are able to support more flexible slicing, we may end up expanding the API of kedro run as well since we can support different combinations of slicing

Possible Implementation

Check out the suggest_resume_scanerio in SequentialRunner:

kedro/kedro/runner/runner.py

Lines 200 to 210 in 91765e3

	remaining_nodes = set(pipeline.nodes) - set(done_nodes)
	postfix = ""
	if done_nodes:
	start_node_names = _find_nodes_to_resume_from(
	pipeline=pipeline,
	unfinished_nodes=remaining_nodes,
	catalog=catalog,
	)
	start_nodes_str = ",".join(sorted(start_node_names))
	postfix += f' --from-nodes "{start_nodes_str}"'

This roughly has the logic to resume pipeline but it's hidden in a private API, we need to surface this for more generic usage.