Concerning the goal to minimize cpu load in generating an excessive amount of thumbnails for a single image, the presented statistics of different thumbnail sizes isn't necessarily meaningful. If each of the images would only ever have exactly one thumbnail generated, it would not harm anyone if the thumbnails all have various and weird sizes. The main problem is that one image has dozens of different thumbnails; so a statistics like "1234 images have 7 thumbnails" would be more meaningful to calculate the possible impact in saving of CPU/storage for additionally generated thumbnails.
Concerning possible sizes of thumbnails, if the space used by them is an issue in any way, I would suggest that there is a predefined list of possible sizes, and an algorithm which takes as input N desired thumbnail sizes, and outputs M thumbnail sizes to be generated/kept, so that at most M thumbnails would be kept.
For the list of those N thumbnail sizes, some suggested ones should be compared with the presented sizes statistics to assess one that matches the best in terms of "minimal amount of scaling needed".