On Proper Learnability between Average- and Worst-case Robustness