Issues with GPU caching

[josiahsams]

The following issues are identified on the Dataset APIs,

1. Cached data should be passed on from one logical plan to another wherever the pointers referring to the columns fields are part of the RDDOUTPUT and not in GPUOUT. Cached GPUOUT fields will continue to be owned by the logical plan which invoke it.

2. Unloading the GPU data should be possible only by the logical plan which owns it.

3. Repetitive loadGpu calls on the same dataset should execute fast as data is already loaded into GPU from the first invocation.