Added Gradio app (and accompanying models and examples). Also added image plots to nn_to_gb notebook.

Added NN features -> LightGBM models, beating all classification-layer-retrained NNs in speed and metric

Split nn training notebook from results and tidied up

Added PyTorch Lightning utilities and model pipeline

Added (Altair) plotting of images and their prediction correctness

Added the utils script since not packaging it yet

First commit: tried 4 NNs for classification, then retrained the smallest one (RexNet 1.0) in full, with unsurprisingly best results by far.