Add support for saving trained models as ONNX

[carterbox]

ONNX is a cross-platform model representation and inference runtime that is used by the NVidia Jetson (for example). While the idea of an open model representation sounds good. There is not official support for converting back to pytorch from ONNX. i.e. You cannot use pytorch as the inferene runtime after saving as ONNX.