Speed up OSQ vector scorer for AVX-512

[iverase]

This commit implements the AVX-512 variant in MemorySegmentES91OSQVectorsScorer which seems to help bulk scoring. The current microbenchmarks looks like:

Benchmark                                                (dims)   Mode  Cnt   Score   Error   Units
OSQScorerBenchmark.scoreFromArray                          1024  thrpt    5   8.336 ± 0.992  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentAllBulk           1024  thrpt    5  15.222 ± 0.351  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentOnlyVector        1024  thrpt    5   8.724 ± 0.092  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentOnlyVectorBulk    1024  thrpt    5   9.444 ± 1.166  ops/ms

With this patch:

Benchmark                                                (dims)   Mode  Cnt   Score   Error   Units
OSQScorerBenchmark.scoreFromArray                          1024  thrpt    5   8.615 ± 0.330  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentAllBulk           1024  thrpt    5  18.649 ± 1.120  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentOnlyVector        1024  thrpt    5   8.633 ± 0.099  ops/ms
OSQScorerBenchmark.scoreFromMemorySegmentOnlyVectorBulk    1024  thrpt    5   9.909 ± 0.097  ops/ms

I have reorganised a bit the code so it is consistent for every method.

[elasticsearchmachine]

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

[john-wagster]

lgtm

Just to state out loud I think we have randomized tests that validate that this produces the same / correct result as before.