Allow limiting the maximum distance for performance

[kno10]

For long sequences, the Levenshtein difference is $O(m*n)$ in complexity.

However, in many cases there will be some small k that gives the maximum distance of interest.

In such cases, the computation can be terminated early if the strings become too different, and if I am not mistaken this allows to reduce the algorithm complexity to $O(max(m,n)\cdot k)$ (basically, of the full m x n matrix you only need to look at a band of height k around the diagonal - pretty straightforward).

[jamesturk]

Makes sense, I'd take a patch for this if there's interest.