ABI handling for more than just specptr/jfptr

[vchuravy]

Due to its heritage GPUCompiler currently only handles specptr and more recently jfptr.

In EnzymeAD/Enzyme.jl#1134 Enzyme is encountering a expectation mismatch. Esentially we ask GPUCompiler to produce a function with specptr ABI and receive a function with ABI japi

using Statistics

function f(temp)
     R_inv = [temp 0. 0. 0. 0. 0.; 
         0. temp 0. 0. 0. 0.; 
         0. 0. temp 0. 0. 0.; 
         0. 0. 0. temp 0. 0.; 
         0. 0. 0. 0. temp 0.; 
     ]

     return R_inv[1,1]
end

To exemplify the issue on 1.10-rc2 hvcat contains an invoke call. So trying to codegen the f function using GPUCompiler yields:

define double @julia_f_340(double %0) local_unnamed_addr {
....
   %47 = call nonnull {}* @ijl_invoke({}* inttoptr (i64 140379636017936 to {}*), {}** nonnull %.sub, i32 31, {}* inttoptr (i64 140379798006736 to {}*))
....

Enzyme then goes and extracts the method instance of the call target (Enzyme free reproducer below) and GPUCompiler produces a

;  @ abstractarray.jl:2114 within `hvcat`
define nonnull {}* @japi1_hvcat_800({}* %function, {}** noalias nocapture noundef readonly %args, i32 %nargs) local_unnamed_addr {

Enzyme most likely will need to handle the multiple ABI's, but I would prefer if there was a better way than just matching the name of the function. Julia internally has the calling convention during codegen and we may be able to put that in meta?

cc: @gbaraldi @wsmoses

Enzyme free reproducer

``` using GPUCompiler using Statistics

function f(temp)
R_inv = [temp 0. 0. 0. 0. 0.;
0. temp 0. 0. 0. 0.;
0. 0. temp 0. 0. 0.;
0. 0. 0. temp 0. 0.;
0. 0. 0. 0. temp 0.;
]

 return R_inv[1,1]

end

module Native
using GPUCompiler
using ReTestItems
setup = include(joinpath(dirname(pathof(GPUCompiler)), "..", "test", "native_testsetup.jl"))
cd(dirname(setup.file)) do
eval(setup.code)
end
end
asm = sprint(io->Native.code_llvm(io, f, (Float64,)))

r_invoke = r"@ijl_invoke({}* inttoptr (i64 (\d*) to {}*), {}** nonnull %.sub, i32 31, {}* inttoptr (i64 (\d*) to {}*))"
m = match(r_invoke, asm)

F = Base.unsafe_pointer_to_objref(Base.reinterpret(Ptr{Cvoid}, parse(UInt, m.captures[1])))
mi = Base.unsafe_pointer_to_objref(Base.reinterpret(Ptr{Cvoid}, parse(UInt, m.captures[2])))

target = Native.NativeCompilerTarget()
params = Native.CompilerParams(false, Native.test_method_table)
config = GPUCompiler.CompilerConfig(target, params; kernel=false, entry_abi=:specfunc, always_inline=false)
job = GPUCompiler.CompilerJob(mi, config)
GPUCompiler.code_llvm(job)

</details>