Transfer of virtuals to the CUDA device now properly uses streams

This way when multiple threads are using VTK-m they all won't block while one transfer a class with virtuals to the device.