ONNX export: add Softmax layer to the end of image classifiers (#57)

When running inference in ONNX, we often want to add a softmax
layer to TorchVision's models.