[FEA] Improve batched `all-neighbors` given CPU indices/distances

[jinsolp]

Current implementation uses managed memory to control single reference of indices/distances.
Using managed memory for large indices/distances arrays may oversubscribe the GPU and lead to performance issues.

Explore performance implications with large indices/distances arrays and work on optimizations if needed.

For example, might need to explore the performance implications of doing a gather->merge on gpu->scatter with the reference indices/distances on CPU memory.