Avx2 decode fix

[lemire]

This is basically PR #30 with various minor fixes.

@Nick-Nuon you had two bugs:

• Instead of Vector256<sbyte> chkVector = Avx2.AddSaturate(Avx2.Shuffle(checkValues.AsByte(), checkHash).AsSByte(), src.AsSByte());, you had Vector256<sbyte> chkVector = Avx2.AddSaturate(Avx2.Shuffle(checkValues.AsByte(), checkHash).AsByte(), src.AsByte());.
• The ToBase64Mask function would return a short, thus truncating the mask (which needs 32 bits).

The AVX results should be 3 times faster at least, but this PR helps a great deal because we no longer fall on the scalar path.

I am not sure I understand what is our performance limitation, but something is hurting us. Still, the results are not bad.

I invite you to run a benchmark.

❤️

[lemire]

@Nick-Nuon I simplified our benchmarks (so that it is faster) and I have added a table to our README. It is not bad.

[Nick-Nuon]

Thanks for the help and pointing these out : it didn't occur to me I might still have bugs.

I get similar results on my rig and otherwise I couldn't find any bugbears so I'm cool with merging. :)

[lemire]

it didn't occur to me I might still have bugs.

To be fair, they were minor bugs. Everything else was fine.

[lemire]

Merging.