Get rid of vertical integrals for IncompressibleModel

[ali-ramadhan]

PencilFFTs.jl and the distributed FFT-based Poisson solver in PR #590 uses dimension 1 to do fast FFTs. Oceananigans.jl uses dimension 3 to do vertical integrals so I don't think we want to decompose the domain in z otherwise we would have to start rewriting kernels.

This leaves us with only the ability to do 1D/slab decompositions in y, which won't scale super well. The "surface to volume ratio" of each rank is quite high leading to large communication overhead: #590 (comment)

Discussing with @glwagner it seems like we should be able to get rid of the vertical integral for the hydrostatic pressure anomaly. It's the only vertical integral left in IncompressibleModel so this would allow for yz/pencil decompositions which should scale much better.

[glwagner]

This isn't pressing for two reasons: (1) #2538 works and (2) we find the quality of solutions degrade when we omit the hydrostatic pressure intergral.