Summary

• add partial Flat-layout random access for fixed-width and ALPRD feature-vector data
• reconstruct directly from selected buffers and avoid repeated expression/metadata work
• batch and share segment-range requests while retaining the default blocking local-file path
• add an in-memory reconstruction benchmark

Why

Uniform feature-vector random access was spending significant CPU in request orchestration, repeated metadata validation, patch handling, and reconstruction. The updated path computes the selected ranges together, shares one exact batch, and reconstructs from the resolved buffers.

This PR is the blocking-I/O control variant. Applying the action/bench-random-access label runs the existing random-access workflow without enabling io_uring.

Measured results

On 8 pinned cores with hot page cache:

• feature-vectors/uniform: approximately 1.95 ms Vortex blocking versus 1.83 ms Lance
• feature-vectors/correlated: approximately 0.34 ms Vortex versus 0.54 ms Lance

Validation

• cargo test -p vortex-array
• cargo test -p vortex-layout
• cargo test -p vortex-io --lib (109 tests)
• cargo clippy -p vortex-io --lib --all-features -- -D warnings
• cargo test --doc -p vortex-io
• release random-access benchmark with 8-core pinning