Support adding CodeInstances to JIT for interpreters defining a codegen cache

[serenity4]

Implements a way to add CodeInstances compiled by external interpreters to JIT, such that they become legal targets for invoke calls.

Based on a design proposed by @Keno, the AbstractInterpreter interface is extended to support providing a codegen cache that is filled during inference for future use with add_codeinsts_to_jit!.

This allows invoke(f, ::CodeInstance, args...) to work on external interpreters, which is currently failing on master (see #57193).

[serenity4]

I also believe that, instead of exposing add_codeinsts_to_jit!, we could define a more generic typeinf_ext_toplevel which would support arbitrary AbstractInterpreters and essentially call add_codeinsts_to_jit after typeinf_ext, as is already done for NativeInterpreter.

@Keno @vtjnash @aviatesk What do you think overall?

[gbaraldi]

I'm a bit concerned about this, specifically because up to now everything that went into the JIT via the usual codepaths used the jl_create_native + JLJITAddLLVMIRModule. But this is not necessarily wrong, also @vchuravy probably has some comments on this

[vtjnash]

This seems reasonable to me, based on the direction Keno has been going with invoke and these other interfaces

[serenity4]

Seems like push!(::IdSet{CodeInstance}, ::CodeInstance) very rarely causes a segfault. I'm not sure what might be causing that given the changes.