Reduce allocations in JaccardSimilarity and JaroWinklerSimilarity

[nishantmehta]

Summary

Two independent allocation reductions in the similarity package:

1. JaccardSimilarity.apply built three HashSets per call: the distinct elements of each input, plus a union set (a copy of the left set with all right elements added) used only to obtain the union size. By the inclusion–exclusion identity |A∪B| = |A| + |B| − |A∩B|, the union set is unnecessary — count the intersection by probing the smaller distinct-element set against the larger and derive the union size arithmetically. The two remaining sets are inherent (the distinct elements of each input) and are now pre-sized to the input length.

2. JaroWinklerSimilarity (matches) materialized two Object[] arrays (the matched characters of each input) solely to count transpositions position-by-position, and tracked matched positions of the shorter input in an int[] where a boolean per position suffices. Count transpositions by walking the matched positions of both inputs in lockstep, comparing characters directly, and use a boolean[] for the match marks.

Behavior is unchanged in both.

Measurement

ThreadMXBean allocation driver, 200k warmed ops, two 19-character inputs:

method	before	after
JaccardSimilarity.apply	2408 B/op	1448 B/op (−40%)
JaroWinklerSimilarity (matches)	376 B/op	144 B/op (−62%)

Testing

JaccardSimilarityTest, JaccardDistanceTest, JaroWinklerSimilarityTest and JaroWinklerDistanceTest pass unchanged (36 tests).

[garydgregory]

Closing. No one's asking for this. Fix bugs in Jira if want to help improve this things.