introduce options for reducing the overhead for a clustering procedure

[alexanderguzhva]

Several changes:

1. Introduce ClusteringParameters::check_input_data_for_NaNs, which may suppress checks for NaN values in the input data
2. Introduce ClusteringParameters::use_faster_subsampling, which uses a newly added SplitMix64-based rng (SplitMix64RandomGenerator) and also may pick duplicate points from the original input dataset. Surprisingly, rand_perm() may involve noticeable non-zero costs for certain scenarios.
3. Negative values for ClusteringParameters::seed initialize internal clustering rng with high-resolution clock each time, making clustering procedure to pick different subsamples each time. I've decided not to use std::random_device in order to avoid possible negative effects.

Useful for future ProductResidualQuantizer improvements.

[mdouze]

Can you say a bit more about when the rand_perm is too slow? It is surprising to me that the rng could be a perf bottleneck.
I assume it's when the number of k = samples needed is << n = total number of ids.
Because the rand_perm is O(n).
When k << n we could sample k elements and checking with an std::set that each elements hasn't been selected before.

[mdouze]

Still importing...

[facebook-github-bot]

@mdouze has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[alexanderguzhva]

I use a dataset 1048576x768 for PRQ experiments with beam_width = 16. In this case with my new candidate code these two mentioned pieces (check for Nans and rand_perm) may take up to 10-15% of the total training time. For rand_perm(), it will be required to generate a permutation of, say, 524288 points out of 16M available.

[facebook-github-bot]

@mnorris11 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[facebook-github-bot]

@mnorris11 merged this pull request in afe9c40.