Convert CUDA tests to use Kokkos

[masterleinad]

This should cover all tests that are not requiring CUDA/GPU/Device-aware MPI, CUDA objects (sparse matrix, vectors, preconditioners), or CUDA MatrixFree.

[masterleinad]

/rebuild

[masterleinad]

Since we are not using any shared memory space like CudaUVM, only the default execution space instance explicitly and only the two-argument deep_copy (that fences for and after all execution space instances), we should be fine without ever fencing explicitly. That being said, I think we should reevaluate the fences after profiling.

[Rombur]

we should be fine without ever fencing explicitly. That being said, I think we should reevaluate the fences after profiling.

Why? If there are not necessary, why do you want to add them?

[masterleinad]

Why? If they are not necessary, why do you want to add them?

I think that it makes sense to at least avoid global fences in certain situations (which we will need to do some more profiling for) and I would rather have the implementation ready for that. If the global fences don't hurt us, adding some more fences is likely also not performance-critical. Either way, I think we should make it easy for users not to forget about synchronization and offer options to be more specific if necessary. Also, I already removed the fences you were commenting on.
I would prefer to get the conversion done and consider synchronization again later.