Dynamic optimization in levenshtein distance

[emilianobovetti]

I tried to write a more efficient implementation for levenshtein distance, forgive me if the code is a bit messy, didn't have time for aesthetics. Anyway benchmarks I ran show appreciable performance benefit:

of course there is still much room for further optimization, but I think it's a starting point!

Since computational complexity is O(n*m) it won't scale well on large strings, I tried on moderately long strings as pointed out in #1 and with following input elm-benchmark gives 92 runs/s on my laptop (i3-6006U).

EditDistance.levenshteinFromStrings
    "abcd111111111100000000001111111111000000000011111111110000000000"
    "111111111100000000001111111111000000000011111111110000000000"

I added fuzz testing just to be sure there aren't uncovered edge cases.