Support ImmersedBoundaryGrid reconstruction using NetCDF

[tomchor]

Another necessary step towards supporting FiedTimeSeries construction from NetCDF output

[tomchor]

One thing that's making this harder is the fact that we started flattening singleton dimensions in NetCDF output. For example, while a 2D Field in Oceananigans may have size(field) == (Nx, Ny, 1), when it is written to NetCDF is has the size (Nx, Ny). In the past this wasn't a problem (and was even desirable to some extent), now when getting information back from NetCDF to Oceananigans, this requires some extra work.

I wonder if it's worth going back on that decision and keeping singleton dimensions. @ali-ramadhan @glwagner thoughts?

[simone-silvestri]

If it makes the source code simpler, I vote for keeping the third dimension. In the end, Oceananigans' fields are always three-dimensional.

[glwagner]

I think it makes sense to keep singleton dimensions in NetCDF output if it simplifies the interface between NetCDF and Oceananigans. I think the disadvantages are minor. If we can identify significant disadvantages in the future, then we can motivate investing energy into supporting the removal of singleton dimensions later.