Struct to string casting functionality

[rwlee]

Resolves #1698
Requires rapidsai/cudf#7406 -- draft until that PR is merged

Inherits incompatibility of base casting -- e.g. floats -- because casting is used recursively

[revans2]

Why do we need to copy this to a column vector? I understand because of the APIs we have right now it is needed, but it would be nice to avoid the copy if possible.

[rwlee]

Ideally we would not copy this column vector, which is why I had been using toString in earlier iterations. How might I avoid this copy?

[revans2]

The signature of doColumnar takes a GpuColumnVector It uses the GpuColumnVector to extract its Spark data type so it can do the cast. A ColumnVector is a ColumnView. As such I would propose that we insert another API in between that we could use for recursive calls.

override def doColumnar(input: GpuColumnVector): ColumnVector =
    doColumnar(input.getBase, input.dataType())

private def doColumnar(input: ColumnView, sparkType: DataType): COlumnVector = {
   (sparkType, dataType) match {
      ...
   }
}

Then you can call the new doColumnar without making copy of the data.

withResource(input.getBase().getChildColumnView(0)) { childView =>
  columns += doColumnar(childView, inputSchema(0).dataType)
}

[rwlee]

There a few things around decimal casting that didn't translate nicely. In some cases they want to do a zero-copy, so I've filtered them out in the GpuColumnVector wrapper API. In those cases, if those cases are found when doing a ColumnView cast, it's copied to a ColumnVector. Reworking the decimal support to be ColumnView friendly version can be done in a follow up PR if necessary, but it seemed a bit out of scope and was non-trivial.

[revans2]

Could we have a follow on issue to clean this up? The issue appears to be because of incRefCount, and there should be ways for us to work around that without too much difficulty.

[rwlee]

#2043

[revans2]

build