Vectorize BQSpaceUtils#transposeHalfByte

[iverase]

After #132797, the method BQVectorUtils#transposeHalfByte can be easily vectorize by using a helper array that defines the shifts we need to apply to the vector elements.

For 128 bits in my mac, the improvements are around 60%:

Benchmark                                           (dims)   Mode  Cnt  Score   Error   Units
TransposeHalfByteBenchmark.transposeHalfByte           384  thrpt    5  3.344 ± 0.100  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte           782  thrpt    5  1.599 ± 0.020  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte          1024  thrpt    5  1.263 ± 0.010  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     384  thrpt    5  1.661 ± 0.031  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     782  thrpt    5  0.805 ± 0.034  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy    1024  thrpt    5  0.625 ± 0.022  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     384  thrpt    5  5.891 ± 0.047  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     782  thrpt    5  2.829 ± 0.025  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama    1024  thrpt    5  2.205 ± 0.042  ops/ms

For 256 bits on a GCP instance shows up to 3x faster:

Benchmark                                           (dims)   Mode  Cnt  Score   Error   Units
TransposeHalfByteBenchmark.transposeHalfByte           384  thrpt    5  1.198 ± 0.035  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte           782  thrpt    5  0.931 ± 0.055  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte          1024  thrpt    5  0.454 ± 0.006  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     384  thrpt    5  0.751 ± 0.033  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     782  thrpt    5  0.347 ± 0.004  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy    1024  thrpt    5  0.284 ± 0.010  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     384  thrpt    5  3.368 ± 0.187  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     782  thrpt    5  1.625 ± 0.004  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama    1024  thrpt    5  1.354 ± 0.059  ops/ms

For 512 bits on a GCP instance up to 2x faster:

Benchmark                                           (dims)   Mode  Cnt  Score   Error   Units
TransposeHalfByteBenchmark.transposeHalfByte           384  thrpt    5  1.251 ± 0.049  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte           782  thrpt    5  0.737 ± 0.013  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte          1024  thrpt    5  0.469 ± 0.036  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     384  thrpt    5  0.683 ± 0.064  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     782  thrpt    5  0.341 ± 0.007  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy    1024  thrpt    5  0.260 ± 0.002  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     384  thrpt    5  2.634 ± 0.056  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama     782  thrpt    5  1.258 ± 0.016  ops/ms
TransposeHalfByteBenchmark.transposeHalfBytePanama    1024  thrpt    5  1.026 ± 0.054  ops/ms

[elasticsearchmachine]

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

[benwtrent]

LOL the panama vector version is easier to read than the regular one !!!!

[ChrisHegarty]

You're on 🔥 @iverase !! LGTM