Use Levenshtein distance instead of Python's difflib

[simonlindholm]

The latter does not obey the triangle inequality -- making a function match worse can improve the score by breaking parts that are accidentally the same between source/target and thus making it line instructions up better.

Using Levenshtein similar to https://github.com/simonlindholm/asm-differ/commit/846a069840a20e099f4b7fb7f1d2b06ea5580b95 would fix this.

Not sure how much of a problem this is in practice.