Re-enable ITime in ocean and sea-ice models

[hgpeterson]

Closes #1992

Following the update to ClimaOcean v0.10 (#1971), stop_time would default to Inf in the ocean/sea-ice models, which forced us to use Floats instead of Dates. This PR in NumericalEarth.jl enables us to set stop_time as a kwarg.

This PR:

• Updates how clock and stop_time are set in the ocean model
• Updates how clock and stop_time are set in the sea-ice model
• Re-enables use_itime: true in config files

This PR also fixes a completely unrelated bug in experiments/test/compare_cmip.jl that emerged from an update in NumericalEarth (see this PR; ocean surface salinity is now passed as a view instead of with interior, changing how we need to define _compare(v1, v2) for SubArrays).

[hgpeterson]

The NumericalEarth PR is merged, but because NumericalEarth is not a direct dependency of ClimaCoupler (it comes in through ClimaOcean), we'll have to wait for a new release there for this PR to work.

[hgpeterson]

An issue that came up that I believe is unrelated to the clock changes: experiments/test/restart_cmip.jl started to fail here with a CUDA scalar indexing error. I believe it has something to do with how the sea-ice model is initialized with SubArrays of OffsetArrays of CuArrays from the ocean (e.g. IceWaterThermalEquilibrium.salinity). To fix the test, I had to add lines of the form

h1 = OffsetArrays.no_offset_view(on_architecture(CPU(), v1))

to experiments/test/compare_cmip.jl, which required adding OffsetArrays to our CMIP Project.toml.

@simone-silvestri Does it make sense to you why separate initializations of the sea-ice model might have different array offsets in their views?

@akshaysridhar Let me know if you're okay with adding OffsetArrays or if we should find a different solution.

[simone-silvestri]

In NumericalEarth/NumericalEarth.jl#346 we have changed the ocean surface salinity to pass
view(S.data, :, :, Nz) instead of interior(S, :, :, Nz) because we need the halos of the surface salinity to compute fluxes inside the halo itself.

https://github.com/NumericalEarth/NumericalEarth.jl/blob/01ab2a6fc1d5e3d6e6b77c20897936495e5c914b/src/Oceans/Oceans.jl#L83-L91

Therefore, the output of that function switched from begin a subarray of a plain array to a subarray of an offsetarray of an array.

I think on_architecture(CPU(), subarray) is enough because it moves everything on the CPU so you avoid the scalar indexing