DL Model Introspection

[dmatveev]

Currently G-API allows to build pipelines of DL models, however the model support is "inference only"

Sometimes, users may need to make decisions based on model information, and G-API doesn't provide such information in any form. E.g., an input layer size to properly adjust preprocessing aspect ratio.

This forces users to do a double-effort - fall back to the original framework, load model manually, and check the layers of interest.

Example: https://github.com/openvinotoolkit/open_model_zoo/blob/master/demos/gaze_estimation_demo/cpp_gapi/main.cpp#L203

G-API needs to provide means to access model information in an abstract, framework-neutral way.