Don't re-infer compiled CodeInstance from external images

[xal-0]

998cb27 introduced a regression that, when invoking a CodeInstance from an external image, would result in the generation of a tojlinvoke trampoline for every invoke but the first (in the order compile! infers them).

module P2

function what(x)
    try
        show(x)
    catch e
        rethrow()
    end
end

what(1) # calls julia_rethrow_88_gfthunk
what(:a) # calls julia_rethrow_88_gfthunk
what('a') # calls jlpkg_rethrow_0

end # module P2

After inferring the first use of rethrow in this example, compile! would find no source for the CodeInstance and infer it again, with source_mode=SOURCE_MODE_GET_SOURCE. It then inserts the resulting new CodeInstance into the code_cache. Before 998cb27, this went into the InternalCodeCache, which would insert it after any CodeInstances with the invoke field set. Now, it is inserted into the OverlayCodeCache, which takes priority over the InternalCodeCache. For the two remaining specializations of what, this new CodeInstance is used for the invoke.

Since it's a waste to re-infer this CodeInstance when we have external_linkage, it makes more sense to thread this through to compile! and skip any CodeInstances that are from the image.

Thanks to @mlechu for helping track down this issue and providing the MWE.

[vtjnash]

IIUC, this function also is supposed to check for the JL_CI_FLAGS_FROM_IMAGE marker. Not required (and generally the cache structure will be expected to surface it first anyways), but defensive?

[IanButterworth]

This seems a more complete (and approved) fix than #62934

So I'll just share the benchmarks I did over there, for visibility, and close #62934

---

This gives a 20% improvement in precompilation time on 1.13 and 10% improvement on nightly

Before
1.13 = #62926 1483760
nightly = ba658eb

With this PR ontop of 1.13 and nightly