Support ZSTD compression method for reading

[stschake]

ZSTD is generally recognized as superior to LZW in size and speed, particularly for decompression.
We use the self-assigned tag first defined by libtiff and also used by GDAL and others.

[fintelia]

Thanks for this PR, zstd support is something I've wanted for a while!

My main concern is actually around build time. On my machine, the zstd crate takes around 10 seconds to compile and this change increases the main image crate's compile time by nearly 25%.

I see two possible options:

• Check whether the ruzstd would work as an alternative.
• Put zstd support behind a feature flag.

[stschake]

Thanks for having a look! I looked into ruzstd but it seems to have a 1.5x-2x performance penalty over the reference implementation, which I felt would diminish the benefit of zstd too much. I added a feature flag now for the zstd support.

[fintelia]

Thanks! If I have a chance, I might look into using ruzstd decoder by default and having the feature opt into the faster zstd bindings. Even the slower decoder may still significantly outperform LZW