Ensembles of simulations for calibrating 3D models

[glwagner]

This PR extends ParameterEstimocean to accommodate "ensembles of simulations". Previously, ensembles were only achievable if the ensemble could be contained within a single Oceananigans.Simulation (ie a (x, y) "forest" of 1D column models in one 3D simulation, or a "cake" of 2D y-sliced models in a 3D simulation). After this PR, it will be possible to set up calibration problems using ensembles of 3D simulations.

There's a simple example in

https://github.com/CliMA/ParameterEstimocean.jl/blob/ensemble-simulations/examples/random_simulation_ensemble.jl

though it isn't literated. That example also illustrates how to implement simple new closures (here a vertical and horizontal tracer diffusivity closure) in Oceananigans, and then calibrate them. The calibration is performed by using just one slice from the 3D model to compare observations and simulation.

[glwagner]

@navidcy we should close this in favor of #295 ?

[navidcy]

Do it!