Reduce heap usage for knn index writers

[benwtrent]

This slightly reduces heap utilization for KnnIndex writers by:

• Only constructing one DocsWithFieldSet instead of 3 for Scalar quantized HNSW formats
• Only having one List<T> instead of 3 for Scalar quantized HNSW formats

This does change an experimental API (FlatVectorFormats) slightly to allow reuse.

[gautamworah96]

Thanks for the quick turnaround on this Ben. I am not too familiar with this code and am taking some time off this week. I'll try to review the code by Friday.

In the meantime, I am also trying to understand some core IndexWriter related concepts. It seems that tuning the indexWriterBufferSizeMB to a lower value seems to be helping (which makes sense because the sum of length of the float[] array across segments being flushed is small, I am trying to see if I can get larger segments by keeping merges enabled, in the email thread I had mentioned that my merges had been disabled).

[benwtrent]

Related: #13553