Improve Reduction Checksums

[MrBurmark]

Summary

Add more accurate kahan and cascade tunings for Base_Seq variants of floating point reduction kernels. This is useful at larger problem sizes where the default naive left fold reduce algorithm becomes inaccurate. This innacuracy can causes GPU variant tunings to spuriously fail their checksums. Adding more accurate reduction techniques like kahan and cascade summation fixes this issue.

Currently the first run variant tuning is used as both the checksum and performance reference. Allow the checksum reference to differ by adding tuning attributes that indicate, among other possible things, which variant tunings have a preferred checksum. Now it will pick the first run variant tuning as the checksum reference but will overwrite that with the first run variant tuning that has a preferred checksum. Note that this has the unfortunate side effect that the -sp output can indicate that variant tuning passed their checksums even if they fail because a later variant tuning may have a preferred checksum, however the output files are correct.

Note that it is a bit strange that I'm calling these Base_Seq variants because they use RAJA classes in their implementations even if they don't use RAJA loop constructs.