[IVF] Improve the format of the tmp file written during merging

[iverase]

During merging, we need to access the vectors in a random access fashion in order to build the clusters. In order to achieve that, we write our vectors and dicIds together on a temporary file. During testing on a memory constraint node, I noticed in the flamegraph that we were taking a lot of time reading docIds:

Looking at this process I noticed we can do much better because:

1. If the segment is dense, e.g all documents have a vector, we don't need to write he docIds as the docId is the ordinal of the vector.
2. If the segment is not dense, we can write the docIds in a separate file as they are access independent of the vectors.

This commit just adds the logic above which improved the performance on memory constraint nodes.

[elasticsearchmachine]

Pinging @elastic/es-search-relevance (Team:Search Relevance)

[benwtrent]

I think separating out the doc ids vs the vectors is great!