[WIP] GPUCompiler + custom JIT

[glou-nes]

Extract from #1754, use a custom GPUCompiler infra / ORC JIT to replace OpaqueClosure.

• Redefine ReactantInterpreter without depending on Enzyme interpreter.
• Incorporate dynamic call rewrite in the type inference and remove fragile custom inference.

I only test this on 1.11, it probably don't run on 1.10. Now, I mostly want to check if that approach can work at scale. I let all debug stuff atm.

[wsmoses]

it still may be useful to have this inherit from the enzyme interp (but we can still specialize the methods below if set to reactant_abi or w/e).

Is it possible/reasonable to do that?

[glou-nes]

Is that possible to elaborate? I don't get what can be used from the enzyme interp.

[wsmoses]

longer term we need to retain the noinline/etc characteristics from the enzyme custom rules, so its preferable to build off the enzyme interp, and perhap specialize it on these methods for the reactant version (e.g. when we have reactant-specific type parameters)

cc @avik-pal

[wsmoses]

can we rename this leak?

also I don't quite follow why this is needed?

[glou-nes]

Sorry, I don't have cleanup my debug stuff. I will remove it.

[wsmoses]

@gbaraldi can you help review?

[wsmoses]

bumping here just because code will get stale over time if not merged (relatedly the 1.12 fixes we did not conflict here)

[glou-nes]

I got three kinds of issues atm. ccall needs the compiler inside some call_with_reactant; type inference changes leading to MethodError and ccall_deferred symbol unknown. It passes half Enzyme Reactant unit tests, and 2/3 of basics.