Vectorize BQVectorUtils#packAsBinary

[iverase]

After #132820, the method BQVectorUtils#packAsBinary 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 almost 2x:

Benchmark                                 (dims)   Mode  Cnt   Score   Error   Units
PackAsBinaryBenchmark.packAsBinary           384  thrpt    5  16.493 ± 0.253  ops/ms
PackAsBinaryBenchmark.packAsBinary           782  thrpt    5   8.193 ± 0.043  ops/ms
PackAsBinaryBenchmark.packAsBinary          1024  thrpt    5   6.461 ± 0.155  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     384  thrpt    5   6.897 ± 0.018  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     782  thrpt    5   1.321 ± 0.007  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy    1024  thrpt    5   2.546 ± 0.043  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     384  thrpt    5  30.626 ± 0.179  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     782  thrpt    5  14.035 ± 0.096  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama    1024  thrpt    5  11.421 ± 0.977  ops/ms

For 256 bits on a GCP instancem looks similar:

Benchmark                                 (dims)   Mode  Cnt   Score   Error   Units
PackAsBinaryBenchmark.packAsBinary           384  thrpt    5   7.620 ± 0.238  ops/ms
PackAsBinaryBenchmark.packAsBinary           782  thrpt    5   3.818 ± 0.306  ops/ms
PackAsBinaryBenchmark.packAsBinary          1024  thrpt    5   3.070 ± 0.025  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     384  thrpt    5   4.321 ± 0.048  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     782  thrpt    5   1.000 ± 0.099  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy    1024  thrpt    5   1.686 ± 0.016  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     384  thrpt    5  13.296 ± 0.270  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     782  thrpt    5   7.064 ± 0.257  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama    1024  thrpt    5   5.941 ± 0.077  ops/ms

For 512 bits on a GCP instance does look so great, still an improvement though. I tried to optimize this case by doing two bytes at a time and shifting the second to build a short but it did not help.

Benchmark                                 (dims)   Mode  Cnt  Score   Error   Units
PackAsBinaryBenchmark.packAsBinary           384  thrpt    5  7.254 ± 0.146  ops/ms
PackAsBinaryBenchmark.packAsBinary           782  thrpt    5  4.057 ± 0.026  ops/ms
PackAsBinaryBenchmark.packAsBinary          1024  thrpt    5  3.213 ± 0.036  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     384  thrpt    5  4.216 ± 0.457  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy     782  thrpt    5  0.892 ± 0.030  ops/ms
PackAsBinaryBenchmark.packAsBinaryLegacy    1024  thrpt    5  1.665 ± 0.107  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     384  thrpt    5  8.862 ± 0.248  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama     782  thrpt    5  4.861 ± 0.268  ops/ms
PackAsBinaryBenchmark.packAsBinaryPanama    1024  thrpt    5  4.037 ± 0.242  ops/ms

relates #132761

[elasticsearchmachine]

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

[benwtrent]

I like it!

I think the bit shifts values should be a private static final value so we don't need to create a new intvector every time the shift is called. But, it looks good to me :)

[ChrisHegarty]

LGTM