Add pure AVX GFNI mode

[klauspost]

Use GFNI even if not on AVX512.

Tested.

Mostly same speed as AVX512 - better than AVX2:

$ ./benchmark -avx512=false
Benchmarking 1 block(s) of 12 data (K) and 4 parity shards (M), each 873814 bytes using 16 threads. Total 13981024 bytes.

 * Encoded 224374 MiB in 10s. Speed: 22436.27 MiB/s (12+4:873814)
 * Repaired 211160 MiB in 10s. Speed: 21115.21 MiB/s (12+4:873814)
 
$ ./benchmark -avx512=true
Benchmarking 1 block(s) of 12 data (K) and 4 parity shards (M), each 873814 bytes using 16 threads. Total 13981024 bytes.

 * Encoded 296560 MiB in 10s. Speed: 29654.76 MiB/s (12+4:873814)
 * Repaired 212347 MiB in 10s. Speed: 21234.58 MiB/s (12+4:873814)

$ ./benchmark -avx512=false -avx2-gfni=false
Benchmarking 1 block(s) of 12 data (K) and 4 parity shards (M), each 873814 bytes using 16 threads. Total 13981024 bytes.

 * Encoded 140600 MiB in 10s. Speed: 14059.98 MiB/s (12+4:873814)
 * Repaired 147347 MiB in 10.001s. Speed: 14733.59 MiB/s (12+4:873814)
 
$ go build -tags=nopshufb
$ ./benchmark -avx512=false -avx2-gfni=false
Benchmarking 1 block(s) of 12 data (K) and 4 parity shards (M), each 873814 bytes using 16 threads. Total 13981024 bytes.

 * Encoded 41693 MiB in 10.003s. Speed: 4168.21 MiB/s (12+4:873814)
 * Repaired 37600 MiB in 10s. Speed: 3759.95 MiB/s (12+4:873814)
 

[klauspost]

Fails... Doesn't seem like any CI machines exposes GFNI.

[klauspost]

$ go test -no-avx512 -timeout=15m
Using SSE2,AVX2,SSSE3,AVX2+GFNI
PASS
ok      github.com/klauspost/reedsolomon        791.824s

[klauspost]

Also this is AVX and not AVX2