Revisit layered ByteStreamSplit design for miniblock encoders

While experimenting with treating ByteStreamSplit (BSS) as a reusable “layer” that wraps other miniblock encoders (RLE, bitpacking, etc.), we discovered a fundamental flaw in the current architecture: BSS rewrites value bytes into fixed 2^k-sized chunks, but it hands the entire rewritten buffer to the inner encoder without exposing those chunk boundaries. Any inner encoder that applies its own chunking (the norm for RLE/bitpack) will cut the BSS-reordered data at different offsets. On read-back, BSS attempts to invert the transform using the inner chunk sizes, so chunks that start in the middle of BSS’s layout are decoded with the wrong byte streams and silently corrupt the values.

In short, “BSS → RLE/bitpack” is unsafe today because the two layers disagree on chunk layout, and the current MiniBlockCompressed format does not capture the additional ordering information needed to reconcile them. Fixing this requires a larger design:

• Give miniblock encoders a way to accept and honor externally provided chunk plans (or expose their own plans for upstream transforms).
• Perform length-preserving transforms (BSS, delta, xor, …) per-chunk, using the same chunk boundaries that the downstream encoder will encode.
• Decode in reverse order using the recorded chunk metadata.

Because this change touches core traits (MiniBlockCompressor), multiple encoder implementations, and potentially new helpers for transform layers, we are postponing the effort.

This issue documents the root problem so we can revisit the layered pipeline design later, once we are ready to refactor chunk planning/coordination across transform and compression layers.

Start by reviewing the MiniBlockCompressor traits, the RLE and bitpacking encoders, and the MiniBlockCompressed format described in the issue. Trace how ByteStreamSplit rewrites buffers and how chunk sizes are used during decoding. Done means the layered pipeline has coordinated chunk plans and enough recorded metadata to reverse the transforms without corruption.