Faster knn chunk merge strategy #4637

Proposal:

Currently knn index is fully rebuilt when disk chunks are merged. A faster approach would be to start with the index from the biggest chunk and add vectors from the remaining chunks being merged. That will improve cpu, i/o and/or memory usage (no need to fetch vectors for the bigger chunk) and chunk merge performance. That, in turn, will improve knn search performance until auto-optimize reached the set chunk count target.

Checklist:

^{To be completed by the assignee. Check off tasks that have been completed or are not applicable.}