The implementation was calling PrepareForOutput on the delegate arrays
rather than PrepareForInPlace, do when used with CUDA you did not get
the data on the device.

Also added a regression test to check this.