Byte shuffle before ZStd

[juntyr]

I was wondering if you have ever benchmarked how ByteShuffle + Zstd would perform in comparison to just ZStd in the case where SZ3 falls back to ZStd, since ByteShuffle can improve performance on structured (non-byte) data

[ayzk]

We primarily focus on lossy cases, so haven't test different lossless fallback strategies. I agree that ByteShuffle + Zstd may improve the ratios.