[8.x] Vector rescoring oversamples k instead of num_candidates

[carlosdelest]

It makes more sense to apply rescoring to an oversampled k instead of num_candidates, as rescoring just a fraction of the candidates will be more performant and offer good recall, specially for smaller k sizes compared to number of candidates.

API changes so we use oversample instead of num_candidates_factor:

GET msmarco-v2-bbq/_search
{
    "query": {
        "knn": {
            "field": "emb",
            "query_vector": [...],
            "k": 10,
            "num_candidates": 100,
            "rescore_vector": {
                "overseample": 2.5
            }
        }
    }
}

This will mean rescoring k * oversample from the num_candidates retrieved on each shard, and returning the top k out of them.

Follow up to #116663

We start with 8.x and will backport to main, as this introduces an incompatible API change that we want to land in 8.x first so BwC and rest compat tests pass in main.

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)