Relocatability of deferred_codegen

[vchuravy]

Currently deferred_codegen pushes a Job into a runtime dictionary. That runtime dictionary is only valid during the session.

I am wondering if instead we could use a de-virtualization strategy similar to Base.

@noinline function gpuc_deferred(f, args...)::Ptr end
@noinline function gpuc_lookup(mi, f, args)::Ptr end

Adding an abstract interpretation extension to refine gpuc_deferred -> gpuc_lookup looking up the corresponding mi.

After codegen (since we can't customize that yet) we scan the LLVM IR for gpuc_lookup and codegen the corresponding functions into the same module.

CUDA wants to get a function pointer and then wrap that in a CuDeviceFunction https://github.com/JuliaGPU/CUDA.jl/blob/e9928ca84509d7c686ea7ec413e1ad2d8176b987/src/compiler/execution.jl#L417