Lazily compute similarity score when reuse the old HNSW graph

[zhaih]

Description

In #12050 we added ability to reuse old graph as an initializer to speed up the merge, but even when we are re-inserting the old graph's node, we still need to calculate a similarity score here so that we can pop-out the worst non-diverse node here based on a sorted order.
But since the score is only used for diversity checking purpose, we probably do not even need them in some cases (like when we never reach the level connection limit). So we probably can first insert those nodes without calculating the score, then when we eventually need to pop out a worst node, we can calculate the score, sort the neighbor array and then do the normal "find the worst node" procedure.