Currently, if we have a situation where majotity of vectors are deleted, indexing of additional paylaods becomes more expensive because of graph merge overhead.
Graph merge happens regardless of the actual numebr of additionally indexed points.
This PR fixes that.

How to repro problem:

docker run --rm -it --network=host qdrant/bfb:dev ./bfb -n 500k -d 128 --keywords 1000 --vectors-per-point 2 --hnsw-m 0 --hnsw-payload-m 16 --segments 1

and then

POST /collections/benchmark/points/vectors/delete
{
  "vector": ["0"],
  "filter": {
    "must": []
  }
}

Indexation of removed vectors will take more time then existing ones.