Speed up BQSpaceUtils#transposeHalfByte

[iverase]

Speed up BQSpaceUtils#transposeHalfByte by unrolly the elements of the array 8 at a time and manually packing the bits. The code is a bit less readable but twice at fast:

Benchmark                                           (dims)   Mode  Cnt  Score   Error   Units
TransposeHalfByteBenchmark.transposeHalfByte           384  thrpt    5  3.356 ± 0.499  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte           782  thrpt    5  1.657 ± 0.014  ops/ms
TransposeHalfByteBenchmark.transposeHalfByte          1024  thrpt    5  1.304 ± 0.021  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     384  thrpt    5  1.719 ± 0.049  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy     782  thrpt    5  0.835 ± 0.005  ops/ms
TransposeHalfByteBenchmark.transposeHalfByteLegacy    1024  thrpt    5  0.652 ± 0.002  ops/ms

It should hint a patch forward for vectorization (if this is not already auto-vectorize which I haven't checked).

relates #132761

[elasticsearchmachine]

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

[benwtrent]

yeah, its an eye-sore.

There must be a faster way (so many of the bit shifts are fairly uniform, making me think they could be done in a single action...oh well).

Progress over perfection :). 2x faster is nice