Improve step detection/transparency

[naglepuff]

Currently, to determine the active learning step, we look at the number of jobs of a particular type run on a folder. This means that if a user deletes the contents of the folder, the current step detected will not reset to zero. We should devise a more sophisticated step-checking method.

We should also try to better communicate the current active learning step to the end user.

[manthey]

The superpixel algorithm generates an epoch number based on annotation names in all of the images. It find the annotations with a particular name and from that finds the highest previous epoch number. Switching to this method would be consistent, but can still be odd if things are manually altered.