MueLu: CreatePreconditioner_PDESystem hangs/fails in fused Jacobi under CUDA

Created by: jhux2

I first saw this on waterman, but it's also happening on geminga:

[snip]
Eigenvalue estimate
Calculating max eigenvalue estimate now (max iters = 10)
Prolongator damping factor = 0.68 (1.33 / 1.94)
Fused (I-omega*D^{-1} A)*Ptent
[hang]

2018-10-17 MueLu dashboard Linux-gcc-5.3.0-OPENMPI-1.8.7_RELEASE_KOKKOS-REFACTOR_EXPERIMENTAL_CUDA-8.0.44

MueLu_UnitTestsTpetra_MPI_1 MueLu_UnitTestsTpetra_MPI_4

Blocks: #2674 (closed), #3482

@trilinos/muelu @csiefer2