I don’t think that mixing VTK and VTK-m’s data models in this way is the best approach. For one, VTK-m arrays can be on the GPU side and using them in VTK may cause them to be copied back to CPU memory. Also, VTK algorithms are not designed to use the VTK-m array API and exposing a different array interface would cause a rift between various algorithm implementations. If we want this (and it is a good idea), we should modify VTK array API to handle implicit arrays. There is already the vtkGenericDataArray infrastructure that would support this.
In the future, the use of VTK-m inside VTK algorithms is likely to grow - mainly because we will want GPU support. When that reaches a critical point, many algorithms can switch wholesale to using the VTK-m API for datasets and arrays.