Revert "Improved fill_halo_regions"

[ali-ramadhan]

Reverts #1444

PR #1444 is quite nice but unfortunately I think the new fill_halo_regions!

fill_halo_regions!(c::AbstractArray, fieldbcs, grid, args...; kwargs...)

broke MPI halo filling since it needs dispatch on

fill_halo_regions!(c::AbstractArray, bcs, arch::AbstractMultiArchitecture, grid, c_location, args...)

It might be tricky to fix since you can't tell whether you're on CPU or MultiCPU from inside fill_halo_regions! with just knowledge of the field...

I propose we revert #1444 just for now (since distributed models + tests are broken on master) and immediately open a new PR with the contents of #1444 where we can find a solution. My only idea is to add architecture to Field but perhaps it feels a little bloated...

This failure was not caught by Buildkite since PR #590 was merged after #1444 was opened, and I think Buildkite only runs tests on branches, not pull requests (i.e. it doesn't merge master and run tests).

[ali-ramadhan]

The implicit free surface solver tests are also failing on the main branch: https://buildkite.com/clima/oceananigans/builds/1644 but this seems like it would be a trivial fix.

[glwagner]

:-O

[glwagner]

It seems reasonable to add architecture to AbstractField. I think we actually might need this to make AbstractOperations work generically on distributed systems (in particular, AbstractOperations that depend on Fields or ComputedFields).

[glwagner]

(Because in compute!, we already try to infer architecture from field.data...)

[ali-ramadhan]

Ah cool, so adding architecture to field seems like it might be good overall.

I'll merge and open a new PR with the contents of #1444.