Use AnonymousSpace to avoid instantiating kernels for both CudaSpace and CudaUVMSpace

[mhoemmen]

See my comment here: kokkos/kokkos#1321 (comment)

Kokkos pull request kokkos/kokkos#1321 fixes Kokkos issue kokkos/kokkos#1175 , by adding a new memory space, AnonymousSpace, that claims accessibility to and from all execution spaces.

This means that code that runs on Cuda, does not depend on UVM behavior, and does not need to allocate Views, should no longer need to distinguish between CudaSpace and CudaUVMSpace. This means, in turn, that kokkos-kernels no longer needs to instantiate most Cuda kernels for both memory spaces. It suffices to instantiate for the Cuda execution space.

This should help with #141 and should also reduce build time and library size.

[mhoemmen]

@prwolfe and @rrdrake may be interested in this.