[Feature Request] Allow splitting the created node graph, just before the model runs, so we get the preprocessed data and use them

Describe the feature request

As a modeler, I would like ONNX to provide me with the ability to split the node graph for my scikit-learn models. I see most scikit-learn models are stored as a node graph of preprocessing steps and the last step is the actual model.

Describe scenario use case

I want to calculate the domain of applicability for the input data to the model and I can't somehow split the created node graph cleanly, so I can get the preprocessed data just before they enter the model.

The issue is that this (the model being the last node in the created node graph) is not the case for all the scikit-learn models and I can't safely extract the preprocessed data just before the enter the model node.

It'd be nice if onnx's model was somehow split into preprocessing nodes, model nodes and post-processing nodes.

The issue names no files, tests, or entry points. Start by locating the scikit-learn conversion and graph-partitioning paths in ONNX Runtime, then determine how preprocessing, model, and post-processing boundaries could be represented. Done means callers can reliably obtain the preprocessed data before model execution across supported scikit-learn models.