JacobiPC and SliceJacobiPC for all-at-once-system

[JHopeCollins]

This PR implements two new preconditioners for the all-at-once system. Each one builds an approximation to the all-at-once Jacobian and applies an (approximate) solve.

1. A block Jacobi preconditioner JacobiPC. This forms the preconditioner from the blocks on the diagonal of the all-at-once system. These blocks are real-valued and have no time average approximation (unless you ask for it), but there is no coupling between the timesteps.

2. A "slice Jacobi" preconditioner SliceJacobiPC. This forms a preconditioner from blocks of several timesteps of the all-at-once Jacobian. The timesteps are coupled within each block, but not between blocks. Another all-at-once preconditioner can be applied to each of these blocks, for example applying CirculantPC to smaller slices of the full timeseries. In this case, the circulant preconditioner in each slice would use the time average from the local slice which would hopefully reduce this error source.

Other additions in this PR

• Base classes for all-at-once preconditioners.
• Renaming DiagFFTPC -> CirculantPC
• A LinearSolver class that solves a ksp with the AllAtOnceJacobian. This does not build the rhs including the initial condition like the AllAtOnceSolver does so unless you provide it with an appropriate rhs it will not solve for an all-at-once-system timeseries.
• A split_ensemble function for taking a large Ensemble and creating smaller sub Ensembles defined over only part of the large one.