Updates for kokkos tuning with multiple ranks

[khuck]

Several issues have come up with respect to tuning kokkos:

• Dealing with multiple ranks:
  • seeding random values in search algorithms with the process rank
  • communicating "best values so far" at periodic intervals (asynchronously)
  • writing the final cache values - one file or an aggregate?
  • Agreeing on common values for contexts, context variables, input variables, output variables
• Handling divergence in solvers (probably with maxiter values)
• doing selective tuning - by name? what about the same name from different code locations?

In addition, in order to handle the Albany land ice case, we have to build a new meta-smoother that sets up and executes the search using different solvers.

Things to consider:

• Hashing strategy for variables - just the name?
• What about repeating variables from different code locations? i.e. "ViewCopy-2D_tile_size_0" will be in any view copy location. Do we want unique variables for each context? Probably, it's easier.
• communication library. Which one to use? zeromq?