Use linear search when small number of candidates to search in

[Kerollmops]

This PR is a proposal akin to meilisearch/arroy#150 to define a threshold at which Hannoy will switch to a linear search instead of using the HNSW. This threshold is arbitrarily set to 1000, but maybe @irevoire or @ManyTheFish has a better idea (5% of the index size)?

We should maybe implement this:
https://github.com/meilisearch/arroy/blob/ee8388229f5c8ec6e3c74c7687061f3e5af92b5f/src/reader.rs#L292-L296

[irevoire]

FYI @Kerollmops I decided to go with 5% of the dataset after looking at your benchmarks here: https://blog.kerollmops.com/meilisearch-vs-qdrant-tradeoffs-strengths-and-weaknesses
Where we see the query time going from 60ms with 6% of the dataset included to 100ms with 1% of the dataset.
I have no idea if hannoy has the same issue and at which thresholds (even for arroy maybe it depends on the number of dimensions or something)

[Kerollmops]

Thank you @irevoire. So, we could keep the current 10000 setting and adjust in the future. @nnethercott we can merge this way I think 😇

[nnethercott]

@Kerollmops @irevoire just ran it on the vector benchmark and it is indeed faster! looks good to me