fix #57749, model ccall binding access in inference

[JeffBezanson]

I've played around with ccall semantics a bunch (#57931) but concluded that a bandaid like this is far better for 1.12.

fixes #57749

[JeffBezanson]

So we basically need a custom interpreter here to handle the non-standard IR in ccall names, which I have tried to do but am flailing a bit 😄 It doesn't really need to be "correct" and its value is not used. It just needs to be enough for inference to tell which bindings are accessed.

So far I only have special support for nests of calls, which handles many cases like A.B.C but I wonder if there are any other syntaxes used out there?

[Keno]

One inline comment and I think we need to pkgeval this to see if there's any corner cases, but I think as a workaround, this seems reasonable.

[JeffBezanson]

@nanosoldier runtests()

[Keno]

@KristofferC already pkgeval'd this in #58987

[JeffBezanson]

This PR wasn't linked nor the tag removed. In any case that run was not too useful since there were assertion failures from the other PR.