Why does BERT has a limitation of only allowing the maximum length of the input tokens as 512?

It's an arbitrary value. It is the longest length of input vector they assumed to be possible. Presumably, they didn't have longer vectors in the training set. Moreover, you can always truncate a vector and ignore farther away history, so in such case the length of the vector would be the farthest history you would considered to be useful. 512 is a power of two, what also suggests that the value is chosen arbitrarily by a computer science minded person.