“For Cursor's codebase indexing system, the primary cost bottleneck is the computation for embedding code, not the storage of vectors in the database.”