This PR adds a getter for class probabilities to segNet.
This getter is zero-copy on a Jetson since all it does is return a pointer to the memory area containing non-argmaxed raw output.
I noticed that segNet::Mask() and segNet::Overlay() do not argmax on GPU but on CPU, so there is no performance loss on performing argmax yourself and this can be useful if you want to do e.g. bilinear filtering on probabilities in order to get less "blocky" results. This is my use case, so I figured somebody else could use this.