Refactor HNSW graph build such that concurrent build won't impact single thread build

### Description

follow-up from #12660, in that PR I made it default that we're going to use Lock and Atomics in both single thread and multithread build. As @benwtrent pointed out it might be better to have separate `NeighbourArray/HnswGraphBuilder` implementation such that we'll separate the single thread from multithread execution and don't add the overhead to single thread build.

Start by reading the referenced PR #12660 and the HNSW graph-building code around NeighbourArray and HnswGraphBuilder. Trace how single-threaded and concurrent builds currently share Lock and Atomics, then define separate implementations whose behavior preserves both build paths while removing unnecessary single-thread overhead.

Written by the indexing model from the issue text.