Only blend with transpose mask in Levenshtein functions if cheaper

[kristoisberg]

Fixes #17, where Restricted Damerau-Levenshtein distances were calculated incorrectly: transposition masks were used even when other operations were cheaper, resulting in higher-than-correct distances. The fix is applied to both distance calculation (including traces) and search functions.

I have to make a disclaimer that while I tested this fix in multiple ways:

• all tests pass, including new test cases based on Wrong result for the case of SIMD enabled (rdamerau) #17 and my own observations
• there isn't a perceivable performance hit in benchmarks
• I tested it against a large set of names and couldn't find any anomalies compared to a known good implementation of regular Levenshtein (PostgreSQL fuzzystrmatch), only expected differences

...the fix itself is mostly AI-generated, as SIMD Rust is quite far out of my comfort zone, so approach with caution. 😄