Raise reconcile_state! kernels to fix GPU segfault

[xkykai]

Closes #407.

On the Reactant GPU backend, reconcile_state!'s plain @jit lowers the coupled model's kernel launch through EnzymeXLA's kernel-call ABI rewrite (rewriteKernelCallABI → gpu.LaunchFuncOp), which segfaults on the large coupled update_state! kernel. This happens at model construction, before any physics runs.

This compiles reconcile_state! with raise=true, so the kernels are raised to StableHLO instead of lowered to a GPU launch, avoiding the crashing path. It's also the lowering the differentiable examples already use for the gradient, so the forward and adjoint passes share the same path.

 function reconcile_state!(model::ReactantESM)
-    @jit Oceananigans.initialize!(model.interfaces.exchanger, model)
-    @jit Oceananigans.TimeSteppers.update_state!(model)
+    @jit raise=true Oceananigans.initialize!(model.interfaces.exchanger, model)
+    @jit raise=true Oceananigans.TimeSteppers.update_state!(model)
     return nothing
 end

Verification (NVIDIA Tesla T4, Reactant 0.2.268)

• The MWE in Coupled model segfaults on the Reactant GPU backend in reconcile_state! #407 no longer segfaults; the coupled model builds on GPU.
• The full differentiable dry-layer slab example runs end-to-end on GPU (build + reverse-mode gradient of the 841-step loop): ∂T/∂𝒫̇ = -1.02e3 K/(kg m⁻² s⁻¹) ≈ -0.047 K/mm, matching the CPU-documented ~0.04 K/mm sign and magnitude.
• Runtime is unaffected (XLA re-vectorizes the raised StableHLO); the only cost is a somewhat longer compile.

🤖 Generated with Claude Code

@dkytezab @glwagner

[codecov]

Codecov Report

❌ Patch coverage is 0% with 2 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
ext/NumericalEarthReactantExt.jl	0.00%	2 Missing ⚠️

📢 Thoughts on this report? Let us know!

[dkytezab]

I this is a reasonably fine workaround for the moment, but we should find the underlying cause of the segfault and address it there.

[xkykai]

ok, will close this once tests pass as a temporary workaround.
I've added a TODO so if the resolution of EnzymeAD/Reactant.jl#3026 solves this problem we can roll back the fix then.

[xkykai]

for the record tests fail right now because of EnzymeAD/Reactant.jl#3028