What the graph really seems to indicate is that time is only linear when working within a cache size on the author's computer (remember that iterating a linked list accounts for the gradual increase in between the cache jumps). If the theoretical upper bound of RAM access was really the important factor at this scale, I wouldn't expect it to be almost flat and to suddenly jerk up every time we have to go to the next cache.

Assuming the author's O(sqrt(n)) is correct, it seems only relevant on much, much larger scales.

In light of that, it really doesn't make sense to pollute the typical use of Big O notation. It should always be understood to be just one metric to understand an algorithm.