After merging #8303 we got a regression on zstd-encoded arrays: #8310 (comment).

There are two issues with that:

• You can't push zstd-encoded byte_length as it decompresses.
• You need to reuse the cache, as branching over ZSTD doesn't solve the issue you recompute values over and over again for each Dict request.

The solution is to use ZstdBuffers and implement a CompareKernel and ByteLength Kernel for it.