MueLu: uncoupled aggregation phase 3 failing on CUDA builds

Created by: lucbv

@trilinos/muelu @rstumin

Expectations

All tests should pass on the CUDA build especially since they might be used by the auto-tester...

Current Behavior

The two CUDA builds with KOKKOS_REFACTOR=ON on geminga are failing due to a failure to aggregates all nodes. This seem like a bug newly introduced in phase 3 of uncoupled aggregation.

Motivation and Context

Mainly we want to maintain a clean dashboard.

Definition of Done

The CUDA builds with KOKKOS_REFACTOR=ON are clean and no tests fail

Possible Solution

I am not sure but it looks like a logic error so probably looking at aggregation phase 3 will be enough