[FEA] Implement a better JNI function to assemble the output columns from cudf::read_json

[ttnghia]

After reading data using cudf::read_json, if the read schema is given, we need to rearrange the output columns from the output table such that the final output columns will have order matched with the column order given in the input schema.

Currently, this process can lead to copying a lot of columns from the output table of cudf::read_json (hundreds column) into a structs column, which leads to significant overhead. We can do much better by just moving them instead, so there will be no data copying at all.