Tpetra CudaUVM flags for when Kokkos_ENABLE_CUDA_UVM goes away

[mcarlson801]

As per the comments on this Tpetra PR, trilinos/Trilinos#12622, our CudaUVM builds should include the following flags to ensure Tpetra allocates in the new Kokkos shared space. These flags will be necessary for when the Kokkos_ENABLE_CUDA_UVM goes away.

Currently, we still need to set Kokkos_ENABLE_CUDA_UVM to on for UVM-on builds since there are some changes still needed in other Trilinos packages. For reference, building with these shared memory allocation flags without the Kokkos_ENABLE_CUDA_UVM will result in errors like these:

/home/maxcarl/weaver/uvm-free-workspace/repos/trilinos-uvmfree/packages/muelu/src/Interface/MueLu_HierarchyManager.hpp(467): error: pointer to incomplete class type is not allowed
          detected during:
            instantiation of "void MueLu::HierarchyManager<Scalar, LocalOrdinal, GlobalOrdinal, Node>::WriteDataAggregates(MueLu::HierarchyManager<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy &, const Teuchos::Array<int> &, const std::__cxx11::string &) const [with Scalar=double, LocalOrdinal=int, GlobalOrdinal=longlong, Node=Tpetra_KokkosCompat_KokkosCudaWrapperNode]"
(306): here
            instantiation of "void MueLu::HierarchyManager<Scalar, LocalOrdinal, GlobalOrdinal, Node>::SetupHierarchy(MueLu::HierarchyManager<Scalar, LocalOrdinal, GlobalOrdinal, Node>::Hierarchy &) const [with Scalar=double, LocalOrdinal=int, GlobalOrdinal=longlong, Node=Tpetra_KokkosCompat_KokkosCudaWrapperNode]"

Opening this issue so that we don't forget to update scripts once the necessary changes are made and we can remove Kokkos_ENABLE_CUDA_UVM.