Pass device memory to MParT through bindings

[mparno]

Ultimately I'd like to create a pytorch Tensor (or Julia array), send it to the device, and then wrap that GPU-memory in a way that Kokkos can use inside MParT. This issue is devoted to assessing the feasibility of that and possibly an initial implementation in the python bindings.

[dannys4]

To build on this/make a suggestion, I'd like to just have an agnostic wrapper for a Kokkos::View that's like MParTArray or something, which we can (in the bindings) create a toBinding or something to create the appropriate array in Python/Julia, as well as toHost, so that passing this in allows us to evaluate maps in the bindings without copying to or from the device