Combine reduction implementations for new and old interfaces

[MrBurmark]

Currently we have a copy of the reduction implementation for the new and old interfaces for at least the GPU. We should be able to combine them into a single implementation. I had this in mind when I implemented the multi-reductions so I think we can probably do something similar and then use that to implement the new reduction interface as well.

[rhornung67]

We have both new reductions and "traditional" reductions available as tunings in RAJAPerf for CUDA and HIP variants. We have not done extensive performance comparisons, which we should do.

Also, consider how to best support users of back-ends for which we only want the new reduction interface to be used, such as SYCL and OpenMP target. For example:

• OpenMP target reductions using the "traditional" interface are slow and do not yield correct results -- they have been turned off in CI testing and removed from RAJAPerf.
• SYCL reductions using the "traditional" interface work except for loc reductions, which are turned off in CI testing. SYCL reductions using "traditional" interface have been removed from RAJAPerf.