Add vector search benchmarks to benchmarking suite

[connortsui20]

Summary

Tracking issue: #7297

Adds a vector-search-bench crate similar to VectorDBBench.

The benchmark brute-forces cosine similarity search over public VectorDBBench datasets (Cohere, OpenAI, etc).

Since Vortex is not a database, we do not measure things like vector inserts and deletes. instead we just measure storage size and throughput for four targets:

• Hand-rolled Rust baseline over &[f32]
• Uncompressed (canonical) Vortex
• Default compressed (which ends up being ALPrd)
• TurboQuant (plus Recall@10 for the lossy TurboQuant path).

Every variant goes through a correctness check against the uncompressed scan before timing.

Not sure if this makes sense to have on a PR benchmark yet since this can only happen on a very specific array tree that will have a very specific optimized implementation.

Testing

N/A

[codspeed-hq]

Merging this PR will degrade performance by 14.69%

⚡ 1 improved benchmark
❌ 1 regressed benchmark
✅ 1151 untouched benchmarks
🆕 10 new benchmarks
⏩ 1455 skipped benchmarks¹

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Mode	Benchmark	BASE	HEAD	Efficiency
⚡	Simulation	old_alp_prim_test_between[f32, 32768]	269.3 µs	223.2 µs	+20.66%
❌	Simulation	new_alp_prim_test_between[f64, 16384]	127.7 µs	149.7 µs	-14.69%
🆕	Simulation	turboquant_decompress_dim128_4bit	N/A	5.5 ms	N/A
🆕	Simulation	turboquant_decompress_dim1024_8bit	N/A	43.9 ms	N/A
🆕	Simulation	turboquant_decompress_dim768_4bit	N/A	42.1 ms	N/A
🆕	Simulation	turboquant_compress_dim128_4bit	N/A	6.6 ms	N/A
🆕	Simulation	turboquant_decompress_dim1024_4bit	N/A	43.9 ms	N/A
🆕	Simulation	turboquant_compress_dim1024_2bit	N/A	47.5 ms	N/A
🆕	Simulation	turboquant_decompress_dim1024_2bit	N/A	43.9 ms	N/A
🆕	Simulation	turboquant_compress_dim768_4bit	N/A	52.1 ms	N/A
🆕	Simulation	turboquant_compress_dim1024_4bit	N/A	52.6 ms	N/A
🆕	Simulation	turboquant_compress_dim1024_8bit	N/A	62.8 ms	N/A

---

_{Comparing ct/vector-bench (b6ea352) with develop (4a5b7d7)}

Footnotes

1. 1455 benchmarks were skipped, so the baseline results were used instead. If they were deleted from the codebase, click here and archive them to remove them from the performance reports. ↩