Add viewer for clean birds and bicycles dataset

[nottombrown]

Currently we save each image into their own files. It would be nice to run a local webserver that renders them all into a single page so that people can inspect them and verify that they are all unambiguous.

I have a messy prototype that I set up for SVHN (adapted from a colab notebook)

https://gist.github.com/nottombrown/783b5553823f4c1b35395fdc144b96fa

Rough spec

• In the evaluate_two_class_unambiguous_model function, dump a map from image_id => (logits, confidence, labels) to a .json file in the plot_confident_error_rate directory
• Set up a minimal flask or node server that parses the json file
• Set up the server to render the viewer using svelte (this should work out of the box)
• Add installation instructions