cuvs-java: Support providing indexing data off-heap

[ChrisHegarty]

Index builders allow to set the input vector dataset (to index) as a float[][]. This requires that all vectors be reified in the Java heap as a single contiguous block of memory. This can be expensive and may exhaust the java heap. Usage example:

float[][] dataset = ...

CagraIndex.newBuilder(resources)
  .withDataset(dataset)

For the Lucene use case, we write the raw vectors to disk and mmap them. To index them, then we need to copy them back into that Java heap as a float[][]. This is not ideal. It would be better to allow the index builder to set the dataset as an opaque pointer + length.