optimize bit unpack on arm64 using neon instructions

[achille-roussel]

Follow up to #3, for bit width 1, 2, 4, and 8, the code now uses vectorized algorithms with NEON instructions to yield much higher throughput:

goos: darwin
goarch: arm64
pkg: github.com/parquet-go/bitpack
cpu: Apple M2 Pro
                       │ /tmp/bench_purego.txt │        /tmp/bench_scalar.txt        │        /tmp/bench_vector.txt        │
                       │        sec/op         │   sec/op     vs base                │   sec/op     vs base                │
UnpackInt32/bitWidth=1            124.40n ± 6%   24.11n ± 4%  -80.62% (p=0.000 n=10)   12.16n ± 4%  -90.23% (p=0.000 n=10)
UnpackInt32/bitWidth=2            124.60n ± 2%   25.89n ± 1%  -79.23% (p=0.000 n=10)   11.77n ± 0%  -90.55% (p=0.000 n=10)
UnpackInt32/bitWidth=4            124.05n ± 0%   26.38n ± 2%  -78.73% (p=0.000 n=10)   11.11n ± 1%  -91.05% (p=0.000 n=10)
UnpackInt32/bitWidth=8            124.25n ± 1%   25.47n ± 2%  -79.50% (p=0.000 n=10)   12.40n ± 3%  -90.02% (p=0.000 n=10)
UnpackInt64/bitWidth=1            125.80n ± 2%   24.45n ± 1%  -80.56% (p=0.000 n=10)   17.88n ± 1%  -85.79% (p=0.000 n=10)
UnpackInt64/bitWidth=2            124.45n ± 5%   24.54n ± 3%  -80.28% (p=0.000 n=10)   20.90n ± 3%  -83.21% (p=0.000 n=10)
UnpackInt64/bitWidth=4            123.85n ± 0%   24.54n ± 2%  -80.19% (p=0.000 n=10)   20.38n ± 0%  -83.54% (p=0.000 n=10)
UnpackInt64/bitWidth=8            124.75n ± 2%   24.61n ± 1%  -80.28% (p=0.000 n=10)   17.47n ± 2%  -86.00% (p=0.000 n=10)
geomean                            124.5n        24.99n       -79.93%                  15.04n       -87.92%

                       │ /tmp/bench_purego.txt │         /tmp/bench_scalar.txt          │          /tmp/bench_vector.txt          │
                       │          B/s          │      B/s       vs base                 │      B/s       vs base                  │
UnpackInt32/bitWidth=1            3.834Gi ± 6%   19.777Gi ± 4%  +415.88% (p=0.000 n=10)   39.230Gi ± 4%   +923.29% (p=0.000 n=10)
UnpackInt32/bitWidth=2            3.828Gi ± 2%   18.420Gi ± 1%  +381.23% (p=0.000 n=10)   40.513Gi ± 0%   +958.42% (p=0.000 n=10)
UnpackInt32/bitWidth=4            3.843Gi ± 0%   18.070Gi ± 2%  +370.22% (p=0.000 n=10)   42.948Gi ± 1%  +1017.59% (p=0.000 n=10)
UnpackInt32/bitWidth=8            3.837Gi ± 1%   18.722Gi ± 2%  +387.94% (p=0.000 n=10)   38.437Gi ± 3%   +901.76% (p=0.000 n=10)
UnpackInt64/bitWidth=1            3.790Gi ± 2%   19.502Gi ± 1%  +414.53% (p=0.000 n=10)   26.679Gi ± 1%   +603.89% (p=0.000 n=10)
UnpackInt64/bitWidth=2            3.833Gi ± 5%   19.431Gi ± 3%  +407.00% (p=0.000 n=10)   22.816Gi ± 3%   +495.32% (p=0.000 n=10)
UnpackInt64/bitWidth=4            3.852Gi ± 0%   19.432Gi ± 2%  +404.53% (p=0.000 n=10)   23.391Gi ± 0%   +507.33% (p=0.000 n=10)
UnpackInt64/bitWidth=8            3.823Gi ± 2%   19.382Gi ± 1%  +407.01% (p=0.000 n=10)   27.297Gi ± 2%   +614.07% (p=0.000 n=10)
geomean                           3.830Gi         19.08Gi       +398.29%                   31.70Gi        +727.73%

[fpetkovski]

🚀