adapt_structure for AbstractArray boundary conditions invokes parent

[glwagner]

This prevents Field and OffsetArray from being used as boundary conditions (correctly) --- if users try to use Field, they get different behavior on CPU vs GPU.

I'm not sure the motivation for this choice.

Oceananigans.jl/src/BoundaryConditions/boundary_condition.jl

Lines 69 to 70 in e16cdc6

	Adapt.adapt_structure(to, b::BoundaryCondition{C, A}) where {C<:AbstractBoundaryConditionClassification, A<:AbstractArray} =
	BoundaryCondition(C, Adapt.adapt(to, parent(b.condition)))

[glwagner]

It looks like it dates back to ancient times practically before the invention of writing

7f4b78e#diff-ef43fe13b0bf0810d1bd93a5cd3264c3be46ba792948fad6a924a62b281e3286

I'm amazed nobody has caught this bug before.

[navidcy]

It may be the answer to several conundrums. Or simply not many use fields for BCs?

[glwagner]

@simone-silvestri may have tried. Not sure. @yuchenma23?

[yuchenma23]

Hi Greg, I was trying field as boundary condition but with time_interpolated_array Simone defined, which is used to interpolate field as a function of time. Essentially it's still a field boundary condition. It never worked on GPU with the bug that is related with some mismatch between types things. But it did work on CPU.

I then got to work on other stuff and never went back to check the code. Perhaps the bug you fixed could explain the bug I faced.

[glwagner]

The fact that time_interpolated_array never worked on GPU explains why you didn't run into this problem!