Support optional threshold parameter for levenshtein function

## Is your feature request related to a problem or challenge?

DataFusion's `levenshtein` function currently only supports the 2-argument form `levenshtein(str1, str2)`. Apache Spark supports a 3-argument form `levenshtein(str1, str2, threshold)` that returns -1 if the edit distance exceeds the threshold, with early termination for better performance.

This missing feature causes downstream projects like [Apache DataFusion Comet](https://github.com/apache/datafusion-comet) to fall back to Spark for queries using the threshold form (see [Comet #3084](https://github.com/apache/datafusion-comet/issues/3084)).

## Describe the solution you'd like

Add an optional third `Int32` argument to `levenshtein`:

```sql
SELECT levenshtein('kitten', 'sitting', 2); -- -1 (distance 3 > threshold 2)
SELECT levenshtein('kitten', 'sitting', 5); -- 3 (distance 3 <= threshold 5)
SELECT levenshtein('kitten', 'sitting', NULL); -- NULL
```

The threshold variant should use a banded DP algorithm with early termination for better performance when the threshold is small relative to string lengths.

### Additional context

Spark's implementation is in UTF8String.levenshteinDistance, based on Apache Commons Text LevenshteinDistance.limitedCompare.

Start by locating DataFusion's existing levenshtein function entry point and its current two-argument behavior. Compare the requested SQL examples with Spark's UTF8String.levenshteinDistance and Apache Commons Text's limitedCompare, then verify that the third Int32 argument handles NULL, returns -1 above the threshold, and uses early termination.

Written by the indexing model from the issue text.