Introduce an int4 off-heap vector scorer

[iverase]

In our IVF implementation, we are currently scoring centroid quantized using int4 on heap. We know that scoring vectors directly from the Index can yield good speedups. Therefore this commit introduces a new class that scorers vectors quantized using int4 off-heap data structures.

this indeed yields a nice speed up:

Benchmark                                             (dims)   Mode  Cnt   Score   Error   Units
Int4ScorerBenchmark.scoreFromArray                       384  thrpt    5  15.384 ± 0.445  ops/ms
Int4ScorerBenchmark.scoreFromArray                       702  thrpt    5   8.908 ± 0.697  ops/ms
Int4ScorerBenchmark.scoreFromArray                      1024  thrpt    5   7.605 ± 0.149  ops/ms
Int4ScorerBenchmark.scoreFromMemorySegmentOnlyVector     384  thrpt    5  16.463 ± 0.456  ops/ms
Int4ScorerBenchmark.scoreFromMemorySegmentOnlyVector     702  thrpt    5   9.854 ± 0.276  ops/ms
Int4ScorerBenchmark.scoreFromMemorySegmentOnlyVector    1024  thrpt    5   8.701 ± 0.865  ops/ms

[elasticsearchmachine]

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

[benwtrent]

This is a good step one towards bulk scoring!

I wonder if we can be even faster if we did "two pass" approximations e.g. dot-product against the centroids with their higher bits, and then only refined with lower bits if within some acceptable error threshold. This would complicate the centroid query logic as we would need to do multiple passes over the centroids. But since we score all of them right now, it seems pretty simple to do.