UCT/CUDA: Skip CUDA cleanup for destroyed contexts (#11437)

* UCT/CUDA: Skip CUDA cleanup for destroyed contexts

Applications can reset the CUDA primary context before MPI_Finalize. Avoid calling cuStreamDestroy or cuEventDestroy on resources from a context that is no longer valid, since CUDA may warn or crash in that path.

* UCT/CUDA/BASE: Retain primary context for CUDA cleanup

* TEST/GTEST/UCT/CUDA: ensure context is being released in test cleanup

---------

Signed-off-by: Roie Danino <rdanino@nvidia.com>