add PQ training benchmark

[sam-herman]

Overview

In order to gauge for variance in PQ performance over various architectures we need to have a baseline measurement within jVector that captures the relative cost of training.
We need to have the number of vectors as an adjustable parameter which is harder to do on a static dataset without creating skewness in data distribution.

Changes

• Added a PQ training benchmark for random vectors with adjustable vector count
• Added PQ build scorer for index construction test to gauge the index construction performance with exact vs PQ build scorer

Testing

PQ training benchmark:

Benchmark                                                                 (M)  (originalDimension)  (vectorCount)  Mode  Cnt     Score     Error  Units
PQTrainingWithRandomVectorsBenchmark.productQuantizationComputeBenchmark   16                  768         100000  avgt    5  2566.605 ± 150.461  ms/op
PQTrainingWithRandomVectorsBenchmark.productQuantizationComputeBenchmark   32                  768         100000  avgt    5  2964.321 ± 124.850  ms/op
PQTrainingWithRandomVectorsBenchmark.productQuantizationComputeBenchmark   64                  768         100000  avgt    5  4028.526 ± 101.389  ms/op

PQ vs FP distance benchmark:

Benchmark                                           (M)  (dimension)  (queryCount)  (vectorCount)  Mode  Cnt       Score       Error  Units
PQDistanceCalculationBenchmark.distanceCalculation    0          768           100          10000  avgt    5  123151.191 ± 42382.755  us/op
PQDistanceCalculationBenchmark.distanceCalculation   16          768           100          10000  avgt    5   11592.736 ±   454.418  us/op
PQDistanceCalculationBenchmark.distanceCalculation   64          768           100          10000  avgt    5   39660.170 ±   315.126  us/op
PQDistanceCalculationBenchmark.distanceCalculation  192          768           100          10000  avgt    5  127094.323 ±  2948.728  us/op

PQ vs FP Index Construction:

Benchmark                                                    (buildScoreProviderType)  (numBaseVectors)  (originalDimension)  Mode  Cnt      Score      Error  Units
IndexConstructionWithRandomSetBenchmark.buildIndexBenchmark                     Exact            100000                  768  avgt    5  19976.081 ±  804.598  ms/op
IndexConstructionWithRandomSetBenchmark.buildIndexBenchmark                     Exact            100000                 1536  avgt    5  37689.636 ± 1295.435  ms/op
IndexConstructionWithRandomSetBenchmark.buildIndexBenchmark                        PQ            100000                  768  avgt    5  25676.173 ± 2423.729  ms/op
IndexConstructionWithRandomSetBenchmark.buildIndexBenchmark                        PQ            100000                 1536  avgt    5  39913.628 ± 4659.553  ms/op

[marianotepper]

LGTM