Allow external packages to influence global load inlining

[wsmoses]

E.g. the following is required to get past the ref inside MPI.SUM.

Similar analogues may be needed for cuda/etc API's, if handled within Enzyme rather than a high-level rule.

using MPI
using Enzyme
using EnzymeCore

Enzyme.API.printall!(true)
push!(EnzymeCore.GlobalInlinedLoads, MPI.SUM, MPI.DOUBLE)

function foo(x::Vector{Float64})
    MPI.Reduce!(x, MPI.SUM, 0, MPI.COMM_WORLD)
    return nothing
end

MPI.Init()

x = ones(10)
foo(x)
x = ones(10)
dx = zeros(10)
autodiff(Reverse, foo, Duplicated(x, dx))

MPI.Finalize()

see: https://github.com/EnzymeAD/Enzyme/pull/1025/files

[vchuravy]

As discussed offline this is better implemented with a custom rule IMO.