Add a Reactant Extension

[glwagner]

There wiill likely be a few methods that need special treatment.

cc @giordano @wsmoses @simone-silvestri

[navidcy]

what's the difference here?

[glwagner]

Basically Ti is the correct choice here, because it is the type of the variables in InterfaceState (typically a floating point). The eltype of the grid, on the other hand, is more general, because it can either be a traced floating point or a constant floating point. The status of the grid eltype might change in the future (eg maybe we will refine the design so that it is always a floating point type). But regardless of how we decide to treat eltype, I think this line is expressing the correct pattern.

[glwagner]

Ok here is a better explanation:

Ti is the eltype of the arrays. This eltype is more general than just a floating point, because in order to transpile Oceananigans/ClimaOcean with Reactant, the eltype of arrays must be "marked" (also known as "traced") as something that can change while the computation is running. This distinguishes the array eltype from the eltype of variables which do not change during a computation (these we refer to as "constants").

Prior to Reactant we did not distinguish between data that can change vs stay the same. Or actually I would say that we more generally simply allow anything at all to change through extensive use of mutable struct. But with Reactant we come into the position of having to specifically state what is changed and what is not.

We can decide to define eltype(grid) as the "underlying" floating point type of an array eltype. I think this might be a valid choice.

However, we retain generality (and in a sense more modularity in the code) if we propagate the array eltypes locally, rather than attempting to always infer the "global eltype" from the grid. The local inference pattern is actually a better / more modular design. It allows us to take the optimal approach for grid eltype regardless of what kernels look like.

I didn't realize any of this until Reactant came around, so I think this is a welcome improvement to realize that local inference is the correct approach. There are a handful of changes sprinkled into Oceananigans that are similar btw.

[giordano]

I'm working on fixing the issue on Windows.

Edit: Fix for making Reactant loadable on Windows available at EnzymeAD/Reactant.jl#1003, but that's not actually making the test fail here because precompilation failure isn't critical and Windows tests don't test Reactant, so all good here.

[glwagner]

I'm gonna merge this, then move the reactant extension tests to github actions where I think they have a better chance at succeeding