Fix policy ordering for CUDA and HIP policies

[michaelmckinsey1]

We noticed some of the CUDA and HIP policies appear to be identical, despite attempting to define different layouts. It would appear that all of these layouts currently execute the GZD policy.

• Should only the ArgumentId nesting match the ordering of the sequential policy or does the ExecPolicy nesting also need to change?
• I noticed a similar looking issue for LPlusTimes.h, which I have not corrected here, and I have not checked the policies in other files.

[pearce8]

@rchen20 @MrBurmark Can you help figure out what these policies should look like? There is definitely a copy paste error making all policies behave the same, but @michaelmckinsey1 wasn't sure he is fixing this correctly. This also seems to be different than the kernel in RAJAPerf.

[MrBurmark]

This does look like a copy paste that was never fixed up. Looking at other policies I can see that the layout is supposed to give the order of the loops, but that doesn't include the moment loop. Due to the tensor contraction nature of the kernel I would think that the moment loop always has to use a sequential policy for correctness. @rchen20 do you know how these are supposed to map to the GPU?

[rchen20]

I am not sure . . . let me show this to John Loffeld, he would know.

[michaelmckinsey1]

these changes are for testing

[rchen20]

After discussion with John, there was a correct ordering some time before we added the HIP backend, and the error likely occurred as HIP policies were added. When I have some time, I'll go through commits and try to find the right one.