Think about how to handle rarely used tunings

[MrBurmark]

Sometimes we want to be able to have multiple different tunings.
However we probably only want to run most of these tunings in rare circumstances like tuning for a new platform.
Examples:

• multiple base implementations ([Draft] Refactor GPU reductions and add unsafe atomic tunings #247)
• different RAJA policies (HALOEXCHANGE_FUSED WorkGroup Dispatch Policies  #260)

One idea is to have them all in the code, but have a runtime option to actually run them. Maybe something like --run-performance-tunings.
Another idea would be to have a configure time option.
Another idea is to leave them in PRs.