Add support for vectorized BMM

[jpountz]

Lucene and turbopuffer now do sparse retrieval with an algorithm that can be described a a vectorized block-max MAXSCORE: https://jpountz.github.io/2025/10/11/vectorized-evaluation-of-disjunctive-queries.html.

In my experience, it performs quite advantageously compared with WAND, BMW, MAXSCORE or vanilla BMM. See for instance how lucene-10.3.0-bp (BMM + recursive graph bisection) performs against PISA (BMW + recursive graph bisection) at https://tantivy-search.github.io/bench/. It could be interesting to support it in PISA to be able to include it in comparisons such as https://jmmackenzie.io/pdf/gmm25-sigirap.pdf cc @JMMackenzie .