Do not re-do inference on panning whole-slide

[kheffah]

I'm just writing this here before I forget. Current system works as follows:

1. Predict on-the-fly all superpixels to show heatmap the first time, then
2. Then predict again every time we zoom to a field.

This means that you re-predict the same superpixels multiple times. Every time the panning window changes, the corresponding superpixel labels are predicted on the fly. This issue could eventually be addressed by some sort of extra database table to store predictions. Keep in mind, the active learning system predicts all superpixels in the entire dataset anyways when we click on “gallery” or when it looks for the most uncertain examples to prompt the user to label, so it’s just a matter of storing these predictions in a table and efficiently retrieving them as opposed to re-predicting on-the-fly.