TileUDT is backed by Arrow record

[echeipesh]

The objective of this epic is to create end-to-end prototype the uses of Arrow in DataFrame.
This is a critical milestone before we start introducing an ArrowCubeUDT (or something like that) into RasterFrames. It's not critical for the Arrow schema we use to be optimal.

At the end of this epic we should have a local environment where some operation is done on ArrowCubeTile on JVM, the raster is sent to python where its transformed using python udf as an ndarray. Some kind of benchmark exists that compares with status-quo.