Settle on a design for interoperability between different `cuda::vector` across heterogeneous memory boundaries

We want `cuda::vector` to be usable accross the board and simplify memory handling for our users.

That requires it to work seamlessly across memory boundaries, e.g how do we envision this to work:
```c++
cuda::vector host_vec { .... };
cuda::vector device_vec = host_vec;
```