Add support for variable sized image preprocessing pipelines

[csvance]

A potential use case for the classification node involves cutting out a section of an image over an object detected with depth sensing/LIDAR and passing it to the classification node. This has the potential to identify more objects than is currently possible with DetectNets with better performance. But the images will be of different sizes when they are sent for classification. The preprocess pipeline CUDAPipeIO needs to contain the input image size as well as allocating memory for a declared maximum input image size.