[BUG] cuvs-java: creating a carga index with a single vectors fails with cudaErrorIllegalAddress:an illegal memory access was encountered

Describe the bug
Creating a carga index with a single vectors fails with cudaErrorIllegalAddress:an illegal memory access was encountered

Steps/Code to reproduce bug

// This is a non-realistic trivial dataset, with just a single small
// vector. It demonstrates that trying to build a cagra index with a
// single vector fails. Such a scenario can happen when data is indexed
// and merged across shards, e.g. in Lucene or Elasticsearch.
@Test
public void testIndexAndSearchSingle() throws Throwable {
  float[][] dataset = {
      { 0.74021935f, 0.9209938f }
  };

  try (CuVSResources resources = CuVSResources.create()) {
    CagraIndexParams indexParams = new CagraIndexParams.Builder()
      .withCagraGraphBuildAlgo(CagraGraphBuildAlgo.NN_DESCENT)
      .withGraphDegree(1)
      .withIntermediateGraphDegree(2)
      .withNumWriterThreads(32)
      .withMetric(CuvsDistanceType.L2Expanded)
      .build();
    CagraIndex index = CagraIndex.newBuilder(resources)
      .withDataset(dataset)
      .withIndexParams(indexParams)
      .build();
    index.destroyIndex();
  }
}

Expected behavior

It will simply java clients if they do not have to special case when the dataset contain a single vector.

Environment details (please complete the following information):

• Environment location: [Bare-metal, Docker, Cloud(specify cloud provider)]
  • AWS g4dn.2xlarge, Ubuntu 22.04 + cuda 12.6

set the JAVA_HOME env variable to JDK 22

JAVA_HOME=/home/ubuntu/binaries/jdk-22.0.2/

Additional context

Fails on branch-25.02

Start with the testIndexAndSearchSingle reproduction and the CagraIndex.newBuilder call on branch-25.02, then run the provided single-vector dataset case in the stated CUDA 12.6 environment. Done means creating and destroying the index completes without cudaErrorIllegalAddress.