perf: add vector throughput benchmark

[westonpace]

There is a rust and a python version. The python version is a little simpler and easier to run but the rust version has been helpful to me for profiling purposes.

We already have some regression benchmarks for latency but throughput can sometimes be different.

[westonpace]

Scores on my local desktop:

ivf_pq_throughput/2.0_1threads_cached
                        time:   [502.18 ms 504.09 ms 505.86 ms]
                        thrpt:  [197.68  elem/s 198.38  elem/s 199.13  elem/s]
ivf_pq_throughput/2.0_1threads_nocache
                        time:   [1.1748 s 1.1769 s 1.1795 s]
                        thrpt:  [84.785  elem/s 84.969  elem/s 85.121  elem/s]
ivf_pq_throughput/2.0_16threads_cached
                        time:   [268.23 ms 270.15 ms 271.91 ms]
                        thrpt:  [367.76  elem/s 370.17  elem/s 372.82  elem/s]
ivf_pq_throughput/2.0_16threads_nocache
                        time:   [308.21 ms 312.46 ms 317.15 ms]
                        thrpt:  [315.31  elem/s 320.04  elem/s 324.46  elem/s]
ivf_pq_throughput/2.1_1threads_cached
                        time:   [513.79 ms 516.59 ms 518.99 ms]
                        thrpt:  [192.68  elem/s 193.58  elem/s 194.63  elem/s]
ivf_pq_throughput/2.1_1threads_nocache
                        time:   [1.1830 s 1.1846 s 1.1861 s]
                        thrpt:  [84.307  elem/s 84.419  elem/s 84.534  elem/s]
ivf_pq_throughput/2.1_16threads_cached
                        time:   [273.70 ms 275.06 ms 276.51 ms]
                        thrpt:  [361.65  elem/s 363.56  elem/s 365.37  elem/s]
ivf_pq_throughput/2.1_16threads_nocache
                        time:   [313.85 ms 317.14 ms 320.12 ms]
                        thrpt:  [312.38  elem/s 315.32  elem/s 318.62  elem/s]

[westonpace]

The numbers are "queries per second". These queries are pretty I/O heavy. With nprobes=20, K=50, and refine_factor=10. So 500 IOPS are needed per query. Even so, the NVMe disk I'm testing against should be able to handle up to 500-800K IOPS/s so I'd expect a roofline around 1000-1600 QPS.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!