[DOC] Support billion-scale CAGRA build with conversion to HNSW

While we have methods to scale CAGRA itself to the billion-scale, we run into some challenges when we build a CAGRA graph (specifically out of core) and convert it to an HNSW graph because the CAGRA graph itself does not fit into host memory.

It's challenging to solve this because the graph optimization step assumes fast random access, which doesn't promote efficient use of disk. We're looking for better ways to break apart the optimization algorithm and/or batch the graph construction so that we can support datasets that surpass the amount of available host memory.

This will enable "dynamic indexing" use-cases where users need to build gigantic graphs that can be immediately consistent/fresh and can't afford to use sharding.

cc'ing @tfeher who's been working on solutions tot his.

The issue names no files or tests. Start by locating the out-of-core CAGRA build and the CAGRA-to-HNSW conversion path, then determine how graph optimization accesses host memory. Done means supporting billion-scale conversion when the CAGRA graph does not fit in host memory.