Vector similarity search scan benchmarks - #7499

Summary

Tracking Issue: #7297

Adds a basic vector similarity search benchmark to the Vortex benchmark suite as a binary.

Here is an example of how to run this:

cargo run -p vector-search-bench --release -- --dataset cohere-small-100k --layout single --iterations 100

cargo run -p vector-search-bench --release -- --dataset bioasq-large-10m --layout partitioned-shuffled --iterations 1

The main scan logic is in scan_one_file in benchmarks/vector-search-bench/src/scan.rs, and everything else is just setup for that.

Future Work

• This does not measure recall, but that can come in a followup PR.
• We will want to have a handrolled baseline implementation that is the "theoretical" minimum over uncompressed f32 vectors that we can compare against. We can then add more flavors of compression and quantization in the future.
• No filter pushdown still (doing filter pushdown before similarity search), have to figure out why that isn't working when we eventually benchmark that
• Recall over filtered cosine search

Testing

The benchmark running successfully for all datasets is sufficient.