Return the result of segmentation in the original size by default

[msluszniak]

Currently, the model expects input of specific shape namely 1024x1024. By default it returns the output with this shape which in my opinion should never happen. As a user, I want always this output image to be the same shape as the original one. So the behaviour should utilize always reshape: true. This can be done in two ways. Either always use OpenCV to reshape the output tensor or try to export segmentation that accepts dynamic shape input and output

[msluszniak]

This needs to benchmarking how much overhead is introduced by cv.resize on each class / argmax. We can either:

• Change default to resize by default and keep the option
• Try to export model with dynamic shape (needs exploration if feasible) (recommended if fast)
• Remove resize and always resize (Now I'm not in favour of this one)