[perf] Bottleneck: gfMul

[qnfm]

Run on latest version 7042874

1. go test -bench=. -cpuprofile cpu.out
2. go tool pprof aes-gem.test cpu.out

One can see

(pprof) top
Showing nodes accounting for 14.30s, 99.37% of 14.39s total
Dropped 11 nodes (cum <= 0.07s)
Showing top 10 nodes out of 28
      flat  flat%   sum%        cum   cum%
    14.10s 97.98% 97.98%     14.10s 97.98%  trailofbits.com/aes-gem.gfMul
     0.10s  0.69% 98.68%      0.10s  0.69%  crypto/internal/fips140/aes.encryptBlockAsm
     0.07s  0.49% 99.17%     14.17s 98.47%  trailofbits.com/aes-gem.(*ghash).update (inline)
     0.03s  0.21% 99.37%     14.20s 98.68%  trailofbits.com/aes-gem.(*ghash).updatePadded
         0     0% 99.37%      0.10s  0.69%  crypto/internal/fips140/aes.(*Block).Encrypt
         0     0% 99.37%      0.10s  0.69%  crypto/internal/fips140/aes.encryptBlock
         0     0% 99.37%     13.85s 96.25%  testing.(*B).launch
         0     0% 99.37%      0.52s  3.61%  testing.(*B).run1.func1
         0     0% 99.37%     14.37s 99.86%  testing.(*B).runN
         0     0% 99.37%      7.08s 49.20%  trailofbits.com/aes-gem.(*aesGEM).Open

[qnfm]

I made two attempts:

1. Replaced gfMul with "gitlab.com/yawning/bsaes.git/ghash" -> led to ~130MB throughput
2. Replace gfMul with assemble ported from go internal ghash (ported amd64) -> led to
   Seal4K: ~460–480 MB/s
   Open4K: ~460–515 MB/s
   Seal64K: ~1.05–1.16 GB/s

The reference optimization is in https://github.com/qnfm/aes-gem/tree/acc