Lucene: Dataset might get flushed twice when using CAGRA build

I ran some experiments with vectorsearch-benchmarks and enabled logging with enableIndexWriterInfoStream. I consistently see about twice the flush size for each segment when using CAGRA_SEARCH instead of CAGRA_HNSW. It looks like the CAGRA index and the dataset is flushed to disk using CagraIndex.serialize() (CuVS2510GPUVectorsWriter.java#L257) which sets include_dataset by default (CagraIndexImpl.java#L369. FlatVectorsWriter writes the dataset to the .vec file regardless. CAGRA_HNSW flushes the adjacency list only (LuceneAcceleratedHNSWScalarQuantizedVectorsWriter.java#L206.

We might want to add a cuvs Java interface for CagraIndex.serialize(OutputStream os, boolean includeDataset) and pass serialize(os, /*includeDataset=*/false) to cuvs-lucene in CuVS2510GPUVectorsWriter.writeCagraIndex().

CC @cjnolet