Optimize the matchLenSSE4 asm implementation.

[nigeltao]

It's also simpler, in that it doesn't use SSE any more.

name old speed new speed delta
EncodeDigitsDefault1e4-8 39.1MB/s ± 0% 39.9MB/s ± 0% +1.89% (p=0.016 n=4+5)
EncodeDigitsDefault1e5-8 30.9MB/s ± 0% 31.8MB/s ± 0% +2.92% (p=0.008 n=5+5)
EncodeDigitsDefault1e6-8 29.7MB/s ± 1% 30.6MB/s ± 0% +3.04% (p=0.008 n=5+5)
EncodeDigitsCompress1e4-8 33.3MB/s ± 0% 33.8MB/s ± 0% +1.49% (p=0.008 n=5+5)
EncodeDigitsCompress1e5-8 23.7MB/s ± 0% 24.2MB/s ± 1% +2.23% (p=0.008 n=5+5)
EncodeDigitsCompress1e6-8 22.4MB/s ± 0% 22.9MB/s ± 1% +2.05% (p=0.008 n=5+5)
EncodeTwainDefault1e4-8 37.4MB/s ± 0% 38.1MB/s ± 0% +1.93% (p=0.016 n=4+5)
EncodeTwainDefault1e5-8 39.3MB/s ± 1% 40.4MB/s ± 1% +2.85% (p=0.008 n=5+5)
EncodeTwainDefault1e6-8 39.3MB/s ± 2% 40.4MB/s ± 2% +2.62% (p=0.032 n=5+5)
EncodeTwainCompress1e4-8 29.3MB/s ± 0% 29.5MB/s ± 1% +0.61% (p=0.048 n=5+5)
EncodeTwainCompress1e5-8 18.6MB/s ± 1% 18.9MB/s ± 1% +1.47% (p=0.016 n=5+5)
EncodeTwainCompress1e6-8 17.0MB/s ± 2% 17.2MB/s ± 1% ~ (p=0.151 n=5+5)

[nigeltao]

The same patch would apply to the snappy package, except #54 is the bigger issue.

[klauspost]

Finally got around to testing it thoroughly. No issues found - thanks for the contribution.