In your MSE case, your network really is outputting a value and its notion of noisyness of that input region. However, to truly be a useful uncertainty image, you ought to look what it outputs outside the training space.

In your toy dataset, you included data from -1 to 7. A good uncertainty metric will grow pretty fast bellow -1 and after 7. In other words, it should say "hey, I have no clue over here"