Add checkpointing/restarting with HDF5

[agzimmerman]

Right now, solutions are only written to VTK files. For post-processing, this is only accurate for CG1 bases. For restarting simulations, this is mostly useless.

sapphire.Simulation should have a simplified interface to firedrake.DumbCheckpoint for checkpointing and restarting.

[agzimmerman]

As of right now, the built-in firedrake.DumbCheckpoint has much of the needed functionality for checkpointing with HDF5. In Sapphire, I still want to leverage the structure of the Simulation class to make checkpointing and restarting simulatoins, as well as post-processing simulation results, more convenient.

[agzimmerman]

To instantiate a simulation, the module in simulations and the arguments to its Simulation.__init__ must be specified.

Many of the arguments are keyword arguments that are readily handled by a kwargs dict. These are in turn readily handled by a JSON.

Some arguments are more complicated. Every module in simulations specifies the weak_form and the element, but the mesh, initial_values, and dirichlet_boundary_conditions are left unspecified. For all of these, as well as some less-used arguments such as nullspace, Python code must be written to specify the argument.

Therefore, in order for Sapphire to have the ability to instantiate a simulation from a JSON, certain requirements must be met, and much generality must be hidden. Specifying a simulation with a JSON is straightforward for any of the modules in benchmarks, because much is assumed about the mesh, initial_values, and boundary_conditions. This reveals that there is a special subset of classes based on sapphire.Simulation which currently only involves the modules in benchmarks. Relating to #29, formalizing which simulations can be run from JSON could help better organize and document Sapphire.

[agzimmerman]

The water freezing benchmark will make a nice test for the JSON specification and checkpoint/restart functionality. Rather than the somewhat complicated implementation currently in sapphire/benchmarks.freeze_water_in_cavity, the initial values can be generated by solving the independently implemented heat driven water cavity benchmark.

[agzimmerman]

The JSON idea was attached to this issue a while back and it still might be nice to provide a simplified interface accepting JSON input for fully specified simulation classes; but right now it's not clear that this will be very useful.

What's really needed is the checkpointing/restarting with HDF5.

[agzimmerman]

Firedrake should handle what we need right now: https://www.firedrakeproject.org/checkpointing.html

At a minimum, sapphire.Simulation should have interfaces to firedrake.DumbCheckpoint.store and firedrake.DumbCheckpoint.load.