Add native int4 operation benchmarks; fix JDK22+ test guards

[ldematte]

Summary

• Add native benchmarks (nativeWithNativeSeg, nativeWithHeapSeg) to VectorScorerInt4OperationBenchmark, comparing native libvec int4 dot product against the scalar and Lucene implementations
• Fix int4 benchmark tests to skip on JDK < 22, since the native scorer requires MemorySegment heap segment support (JDK 22+)
• Add test assertions for the new native benchmarks in VectorScorerInt4OperationBenchmarkTests

Cross-check benchmark results (Mac, Apple M2)

These are just a local cross-check. Full benchmark results on cloud instances are in #144429 and #144649.

Benchmark                                               (size)  Mode  Cnt    Score    Error  Units
VectorScorerInt4OperationBenchmark.scalar                 2048  avgt   15  651.558 ±  1.992  ns/op
VectorScorerInt4OperationBenchmark.lucene                 2048  avgt   15  697.716 ± 11.837  ns/op
VectorScorerInt4OperationBenchmark.nativeWithHeapSeg      2048  avgt   15   41.390 ±  0.409  ns/op
VectorScorerInt4OperationBenchmark.nativeWithNativeSeg    2048  avgt   15   42.997 ±  0.162  ns/op

Test plan

• VectorScorerInt4OperationBenchmarkTests passes (new native assertions)
• VectorScorerInt4BenchmarkTests passes (JDK22+ guard)
• VectorScorerInt4BulkBenchmarkTests passes (JDK22+ guard)

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[john-wagster]

lgtm

[john-wagster]

retried serverless checks a couple of times; there's a PR to fix the failing test in flight: https://github.com/elastic/elasticsearch-serverless/pull/6048

merging.