GPUCompiler doesn't support deferred codegen more than second order

I know there's a bunch of misc PR's open for resolving, but I didn't see a tracking issue so opening this as something to point folks to.

x/ref in this repo https://github.com/JuliaGPU/GPUCompiler.jl/pull/625 and https://github.com/JuliaGPU/GPUCompiler.jl/pull/599 and https://github.com/JuliaGPU/GPUCompiler.jl/pull/582 and https://github.com/JuliaGPU/GPUCompiler.jl/pull/556

x/ref https://github.com/JuliaSmoothOptimizers/NLPModels.jl/issues/482 https://github.com/EnzymeAD/Enzyme.jl/issues/1707 https://github.com/EnzymeAD/Enzyme.jl/issues/1173

cc @jgreener @michel2323 @tmigot @vchuravy @Arpit-Babbar @cgeoga

Start by reading the linked GPUCompiler.jl pull requests #625, #599, #582, and #556, then compare the related NLPModels.jl and Enzyme.jl issues. The issue provides no specific files or tests; done would require deferred code generation to work beyond second order.

Written by the indexing model from the issue text.