Add sigmoid_fp64 kernels

[copybara-service]

Add sigmoid_fp64 kernels

And rewrite the sigmoid_fp32 kernel using the same technique.

It turns out that this kernel is faster, which is a little surprising. It does have less "overhead" (special cases, piecewise branches, etc.) in exchange for more polynomial arithmetic.

Change in performance for sigmoid_fp32:

name                                                                time/op        time/op     vs base               
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time   1.759µ ± 3%   1.771µ ± 21%        ~ (p=0.818 n=6)
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time   1.740µ ± 3%   1.774µ ±  3%        ~ (p=0.065 n=6)
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time   1.773µ ± 2%   1.783µ ±  1%        ~ (p=0.699 n=6)
bench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time               4.909µ ± 1%   3.670µ ±  2%  -25.24% (p=0.002 n=6)
bench/sigmoid_fp32_1x16_x86_avx2/m:4/n:1024/real_time               4.830µ ± 2%   3.706µ ±  4%  -23.28% (p=0.002 n=6)
bench/sigmoid_fp32_1x16_x86_avx2/m:16/n:256/real_time               4.912µ ± 1%   3.740µ ±  2%  -23.87% (p=0.002 n=6)
bench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time               6.632µ ± 3%   5.437µ ±  2%  -18.02% (p=0.002 n=6)
bench/sigmoid_fp32_1x32_x86_sse2/m:4/n:1024/real_time               6.637µ ± 3%   5.524µ ±  3%  -16.77% (p=0.002 n=6)
bench/sigmoid_fp32_1x32_x86_sse2/m:16/n:256/real_time               6.692µ ± 4%   5.493µ ±  1%  -17.92% (p=0.002 n=6)
geomean                                                             3.851µ        3.305µ        -14.19%

sigmoid_fp64 compared to other kernels:

----------------------------------------------------------------------------------------------------------------------------
Benchmark                                                                  Time             CPU   Iterations UserCounters...
----------------------------------------------------------------------------------------------------------------------------
bench_reference/sigmoid_float/m:1/n:4096/real_time                     38579 ns        38571 ns         7348 Bytes=849.382M/s Op=106.173M/s
bench_reference/sigmoid_float/m:4/n:1024/real_time                     38345 ns        38338 ns         7440 Bytes=854.556M/s Op=106.819M/s
bench_reference/sigmoid_float/m:16/n:256/real_time                     39192 ns        39187 ns         7190 Bytes=836.095M/s Op=104.512M/s
bench_reference/sigmoid_double/m:1/n:4096/real_time                    91326 ns        91313 ns         3045 Bytes=717.606M/s Op=44.8504M/s
bench_reference/sigmoid_double/m:4/n:1024/real_time                    91307 ns        91290 ns         3043 Bytes=717.757M/s Op=44.8598M/s
bench_reference/sigmoid_double/m:16/n:256/real_time                    93505 ns        93486 ns         3018 Bytes=700.885M/s Op=43.8053M/s
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:1/n:4096/real_time       1786 ns         1786 ns       155658 Bytes=18.3422G/s Op=2.29277G/s
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:4/n:1024/real_time       1802 ns         1802 ns       157599 Bytes=18.1847G/s Op=2.27309G/s
bench/sigmoid_fp32_1x32_x86_avx512f_avx512bw/m:16/n:256/real_time       1791 ns         1791 ns       156134 Bytes=18.2963G/s Op=2.28704G/s
bench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:1/n:4096/real_time       4475 ns         4475 ns        60425 Bytes=14.6433G/s Op=915.207M/s
bench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:4/n:1024/real_time       4822 ns         4821 ns        59593 Bytes=13.5913G/s Op=849.459M/s
bench/sigmoid_fp64_1x16_x86_avx512f_avx512bw/m:16/n:256/real_time       4842 ns         4840 ns        56596 Bytes=13.5363G/s Op=846.016M/s
bench/sigmoid_fp32_1x16_x86_avx2/m:1/n:4096/real_time                   3789 ns         3788 ns        69486 Bytes=8.64752G/s Op=1.08094G/s
bench/sigmoid_fp32_1x16_x86_avx2/m:4/n:1024/real_time                   3892 ns         3892 ns        74142 Bytes=8.41825G/s Op=1.05228G/s
bench/sigmoid_fp32_1x16_x86_avx2/m:16/n:256/real_time                   3757 ns         3756 ns        72827 Bytes=8.72073G/s Op=1.09009G/s
bench/sigmoid_fp64_1x8_x86_avx2/m:1/n:4096/real_time                   10451 ns        10450 ns        26516 Bytes=6.27103G/s Op=391.939M/s
bench/sigmoid_fp64_1x8_x86_avx2/m:4/n:1024/real_time                   11010 ns        11007 ns        24451 Bytes=5.95261G/s Op=372.038M/s
bench/sigmoid_fp64_1x8_x86_avx2/m:16/n:256/real_time                   10475 ns        10472 ns        26374 Bytes=6.2567G/s Op=391.044M/s
bench/sigmoid_fp32_1x32_x86_sse2/m:1/n:4096/real_time                   5649 ns         5648 ns        49675 Bytes=5.80048G/s Op=725.06M/s
bench/sigmoid_fp32_1x32_x86_sse2/m:4/n:1024/real_time                   5646 ns         5645 ns        50916 Bytes=5.80353G/s Op=725.441M/s
bench/sigmoid_fp32_1x32_x86_sse2/m:16/n:256/real_time                   5571 ns         5571 ns        48792 Bytes=5.88151G/s Op=735.188M/s
bench/sigmoid_fp64_1x8_x86_sse2/m:1/n:4096/real_time                   15957 ns        15952 ns        17116 Bytes=4.10712G/s Op=256.695M/s
bench/sigmoid_fp64_1x8_x86_sse2/m:4/n:1024/real_time                   15657 ns        15654 ns        17451 Bytes=4.18581G/s Op=261.613M/s
bench/sigmoid_fp64_1x8_x86_sse2/m:16/n:256/real_time                   15748 ns        15744 ns        17804 Bytes=4.16163G/s Op=260.102M/s