Init HNSW merge with graph containing deleted documents

### Description

Currently when we're merging HNSW graphs we're able to start with an exiting graph but not inserting nodes from scratch thanks to #12050. But we have set a constraint that the init graph cannot be the one that has deletions since we do not have a good way to handle deletions within a graph without reconstructing it. And this makes the nice optimization a bit wasteful as in reality the bigger the segment the less probability it has no deletion, especially in a hybrid text-embedding index.

Opening this issue to discuss how we could solve this problem or have a work around.
My very naive ideas:
1. We could remove the deleted node from the graph, fully connected all its neighbours, and do diverse check on those neighbors to remove extra links. In case the deleted node is an entry node, we can insert the closest node of the deleted node upto where the deleted node existed.
2. We tolerate certain amount of deletions (like 10% ~ 20%) inside HNSW graph and just use them as connections.

I personally like 1 better as for 2 we need to change the searching behavior accordingly and sounds a bit fragile. Any other ideas are welcome!

The issue names no files or tests. Start by reading the HNSW merge work in #12050 and locate the constraint that rejects an initialization graph with deletions. Compare the proposed deletion-handling approaches and define an agreed design and validation plan before implementation.

Written by the indexing model from the issue text.