Abstraction for configuring kernels

[glwagner]

@simone-silvestri's recent work on scalable performance for the HydrostaticFreeSurfaceModel used the concept of "inner" and "boundary" regions to overlap communication and computation.

The basic components are:

1. Reduce communication as much as possible (for example, compute diagnostic variables wherever needed rather than communicating results)
2. Compute the "interior" of the tendencies first, which do not require halo regions, while communication is occurring
3. Once communication is finished, compute the "boundary regions" of the tendency

Point (1) requires us to compute diagnostic variables inside halo regions. This requires both expanded sizes and offsets for each diagnostic variable. Different diagnostic variables have different sizes --- the width of the diagnostics kernels depends on how many points are accessed.

Points (2) and (3) require computations over parts of the grid (reduced size + offset).

These considerations demand a nice abstraction so that the code for launching kernels is readable and understandable (and also concise).

[glwagner]

@simone-silvestri make sure you link to this issue when you open a PR to close it