Keep dataset shape and use shaped data for training and prediction

[juliabeliaeva]

This PR is a continuation of #494. It adds shape information for all the data used for prediction and training. Since shape is always known, we can create tensors for the data more easily.

Some models do not have the input shape defined and data with arbitrary shape can be used (ex EfficientDetObjectDetectionModel). In this PR such models still use resize operation in preprocessing, so they still have some input shape set manually for this. Maybe this is not needed (?)

[ermolenkodev]

Some models do not have the input shape defined and data with arbitrary shape can be used (ex EfficientDetObjectDetectionModel). In this PR such models still use resize operation in preprocessing, so they still have some input shape set manually for this. Maybe this is not needed (?)

I think it's okay to have explicit resize in examples because it's pretty common to explicitly resize the image before inference because size has an effect on performance, moreover, some models work better with specific input sizes, although support variable input.