Implement memory pool allocator

[abullock-lanl]

This adds the infrastructure for a custom GPU-enabled memory pool allocator.

More extensive testing and documentation is needed before I can mark this ready. I will likely also add a process error handler.

[abullock-lanl]

Update: I've added an error handler function for graceful exits that also uses libunwind for stack traces (if available on the given platform).

[abullock-lanl]

Here's some of the next steps I'm thinking about (not for this PR) to fully adopt the new array types here:

• We want to replace the usage of all STL vectors for things that will be used on device to Kokkos views or std::array instead of the current process of creating views from STL vectors. The simplest thing to start with would be replacing local DBLV_T variables with the new ArrayRank1<double> type that gets returned from a memory pool allocation. For instance:

- DBLV_T point_volume(pll, 0.0);
+ auto point_volume = NewArrayRank1<double>("pvol", pll);

• Replacing the STL vector usage for database accesses can come later.
• Before we can replace the simple vectors, the MPI comm machinery has to be updated to operate on views/arrays instead of just STL vectors, otherwise we can't do things like gather/scatters on the point volume example above.
• Should I maintain left layout for memory pool arrays? I think probably not since this can be awkward in the UME context where looping over an inner dimension variable does not happen (we always have 3-element std::array types as the "inner" dimension).