Work-around for ROCM overflowing ints when scan-summing arrays

[gunney1]

Summary

• This PR is a bugfix (work-around)
• It does the following:
  • Use a totally inefficient 32-bit int to represent a boolean, to work around a ROCM behavior.
  • Add comment about the same work-around for another array.
  • Add notes on when we can remove the work-around.

We have 2 arrays, one with booleans and one with ints in [0,5], but we have to use more bits to represent them for ROCM. We do sum scans on them and put the output in 32-bit ints. However, ROCM is doing the sum in the input type without promoting to the output type. This work-around bumps up the input type.

Related to llnl/RAJA#1631

[kennyweiss]

Thanks @gunney1

Clearly, this is necessary.
Other than more memory, are there performance implications?

[kennyweiss]

Looks like you missed an initialization:

/home/axom/axom/src/axom/core/../../axom/quest/detail/MarchingCubesImpl.hpp:77:7: error: invalid initialization of reference of type 'axom::Array<int>&' from expression of type 'axom::Array<long int>'
   77 |     , m_facetIncrs(facetIncrs)
      |       ^~~~~~~~~~~~~~~~~~~~~~~~

https://dev.azure.com/axom/axom/_build/results?buildId=13141&view=logs&j=0b50086d-0dae-5824-d20c-f0d441949587&t=ebe88c80-2dbb-5dd9-2006-9b9be4efc4c3