[FEA][JNI] Leverage cub's multi-buffer copy algorithm in JNI bindings

We would like to use https://github.com/NVIDIA/cub/issues/297 for our spilling logic https://github.com/NVIDIA/spark-rapids/issues/7672 in the spark-rapids plugin.

We currently contiguous_split every single buffer that _could be_ spilled in the future, because at the time of spill the last thing we want to do is double our memory usage right before copying to host. The multi-buffer work should allow us to bounce-buffer out of GPU memory at spill time, without having to contiguous_split for spillable buffers.