Add t-digest 3.2 vs 3.3 comparison and merge-order reproducer

[xiangfu0]

Summary

• add a direct t-digest 3.2 vs 3.3 comparison test on a deterministic Pinot-like hierarchical merge dataset
• keep the pure 3.3 merge-order reproducer that shows compression 500 restoring stable hierarchical merges
• copy com.tdunning:t-digest:3.2 into the test target so both versions can be exercised side-by-side in the same Surefire JVM

Validation

• ./mvnw -pl pinot-segment-local -Dtest=TDigestVersionComparisonTest,TDigestMergeOrderReproducerTest -Dsurefire.failIfNoSpecifiedTests=false test
• same command passed in 10 fresh Surefire JVM runs (10/10)
• ./mvnw -pl pinot-segment-local spotless:apply
• ./mvnw -pl pinot-segment-local checkstyle:check
• ./mvnw -pl pinot-segment-local license:format
• ./mvnw -pl pinot-segment-local license:check

Comparison Signal

On the minimized exact-quantile comparison scenario, the deterministic dataset shows a clear gap:

• 3.2 @ compression 100: about 0.000074 max normalized error with 121 centroids
• 3.3 @ compression 100: about 0.005094 max normalized error with 55 centroids
• 3.3 @ compression 150: about 0.000049 max normalized error with 79 centroids

Context

This PR is intentionally split out from #18103 so the reproducer and the direct 3.2 vs 3.3 comparison can be reviewed independently from the production compression change.