Investigate lazy filter iteration for IVF

[benwtrent]

Description

For HNSW, building the bit-set up-front for filtering is pretty much required as vector ordinals will never be iterated in doc-id order.

For IVF, we do iterate each of the postings list in doc-id order. meaning, we can simply apply the filter as each posting list is scored.

I think to test this, we should:

• have completely random filters
• positively correlated filters (e.g. filtered docs are also the nearest...)
• negatively correlated filters (e.g. filtered docs are the furthest)

It can be costly to iterate ALL docs to gather a filtered bit set and then apply it, this optimization could be quite good for filtered vector search.

//cc @jimczi

[elasticsearchmachine]

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