[DOC] Benchmark matrix across RAFT ANN algorithms

[cjnolet]

We've been asked by several folks for a matrix of benchmarks across different algorithms across different scales of datasets. We should just go ahead and do one which we publish on RAFT's documentation (preferably using RAFT's bench-ann Python scripts to make them consistently reproducible).

I would propose the following matrix:

Batch Size: 1, 100, 10k
Scales: 100k, 10M, 100M
K: 1, 10, 1k
Hardware: T4, V100, A100, H100 (potentially L4, A10)
Algorithms: Brute-force (for smaller scale), IVF-PQ, IVF-Flat, CAGRA

Datasets:

1. For the various data scales, I would propose using MS-turing-1B, which includes a 10M subsampled version in big-ann-benchmarks '23 (https://big-ann-benchmarks.com/neurips23.html) and we can use 100M subset from the whole thing in our bench-ann suite from big-ann-benchmarks '21 (https://big-ann-benchmarks.com/neurips21.html).
2. We are going to want to find a 100M scale dataset with higher dimensions (~512 would be best).

[achirkin]

I would also propose couple more variables to the mix to more easily spot the problems like #1726:

• top-k value k: 1, 10, 20, 50, 100, 200, 500, 1000
• If we end up using a few different datasets, make sure we have at least one of each data type we support (int8/uint8/float for now)

[achirkin]

Also I think this is a good case to advertise my gbench PR #1661, which should make writing the configs much easier and speed-up the benchmarks by orders of magnitude.

[MarkMoTrin]

Would be good to add the L40S as well

[cjnolet]

@achirkin, if we can prioritize getting the gbench changes so the outputs function like the existing benchmarks (the benchmarks dump to files), then I'm definitely on board w/ getting those merged in. I'd like to avoid breaking the Python scripts in the meantime now that we're pointing users to them.