Enable OnPair in GPU compression benchmark

[joseph-isaacs]

Summary

• retain OnPair alongside FSST in the CUDA-compatible BtrBlocks compressor
• avoid forcing Delta for OnPair offset children when that unsupported GPU scheme is excluded
• stage device-resident OnPair dictionary metadata on the host before preparing the CUDA dictionary

The GPU compression workflow and documented local command already build with cuda,unstable_encodings; this makes the CUDA-compatible compressor actually consider OnPair under that feature set.

Local evidence

Command:

cargo run -p compress-bench --profile release_debug \
  --features cuda,unstable_encodings -- \
  --gpu-decompress --iterations 5

String-heavy TPC-H l_comment canonical, based on merged develop after #8920:

Variant	GPU decompress time
merged develop (FSST)	1,114,305 μs
this branch (OnPair selected)	1,069,992 μs
change	-3.98%

I removed the old capture and regenerated /tmp/onpair-gpu-decompress.vortex from this branch. The new file is a valid 222,367,880-byte VTXF file (SHA-256 0bb262282ff5c3fcd3459141cffbc01f6f465c7287221f29d4ffa4853cb396e3). A fresh-file runtime trace confirms actual use, rather than only registry presence:

try GPU dispatch encoding=vortex.onpair dtype=Utf8(NonNullable) len=8192

Checks

• cargo +nightly fmt --all -- --check
• cargo test -p vortex-btrblocks --features unstable_encodings builder::tests::cuda_compatible_uses_gpu_string_encodings
• cargo test -p vortex-cuda onpair -- --nocapture (19 passed)
• cargo clippy -p vortex-compressor -p vortex-btrblocks -p vortex-cuda --all-targets --all-features
• end-to-end GPU file benchmark and traced OnPair CUDA dispatch