Fix higher order codegen

[wsmoses]

No description provided.

[vchuravy]

What's the GPU story here?

[wsmoses]

So the follow up there is basically to also have a similar handler that looks for deferred_(id) and we'll just run gpucompiler codegen there and inline the module similarly to here where we do the mapping variant from the pointer

[github-actions]

Benchmark Results

	main	d629c30...	main/d629c3057119ec...
basics/overhead	4.33 ± 0.01 ns	4.34 ± 0.01 ns	0.998
time_to_load	1.14 ± 0.0098 s	1.2 ± 0.043 s	0.952

Benchmark Plots

A plot of the benchmark results have been uploaded as an artifact to the workflow run for this PR.
Go to "Actions"->"Benchmark a pull request"->[the most recent run]->"Artifacts" (at the bottom).

[vchuravy]

I want to remove the deferred_id buisness from GPUCompiler so that we can actually precompile GPU kernels,
and I am not keen on maintaining two code-paths.

[wsmoses]

Hm how is pre compilation blocked?

but in any case yeah this PR primarily just fixes the “Enzyme.autodiff is the outermost compiler job” context

[wsmoses]

More thought/design is merited for when that is not the case

[vchuravy]

So this PR is the proposed new API for deferred compilation in GPUCompiler JuliaGPU/GPUCompiler.jl#582 this is without any changes for Enzyme.

The rest of the stack is one possible way to add the necessary bits and bobs to make it Enzyme compatible.
Generally speaking, the rule for precompilation is "you shan't depend on runtime data" so pointers or incrementing counters are out.

[vchuravy]

Hm how is pre compilation blocked?

We will serialize a Julia CodeInfo that contains a magical pointer/id? That is then used for a dictionary that only exists during the runtime of a particular code.

[wsmoses]

hm yeah fair enough.

In any case even if not a final resolution, would you be opposed to merging?

This net deletes code (including getting rid of a lot of runtime pointers) and removes the autodiff_deferred upgrade in absint, and thus fixes issues like #2147

[wsmoses]

[to be clear we would still need to follow up/continue here, but it also at least unblocks the CPU side higher order stuff]