[IR] Provide well-formed intermediate in ApplyPassToFunction

[Lunderberg]

Prior to this commit, ApplyPassToFunction removed functions from the IRModule to hide them from the inner ir.transform.Pass. The dangling GlobalVar references to those functions meant that the intermediate IRModule was ill-formed This commit updates the ApplyPassToFunction utility to instead replace the functions with ExternFunc nodes. This still prevents the inner ir.transform.Pass from having visibility into functions that should not be mutated, but provides a well-formed IRModule.

[Lunderberg]

A follow-up commit to #16801

[slyubomirsky]

Are the substituted global vars restored afterwards? I couldn't find a place where they were put back. I would be worried about these extern funcs ending up in the final module.

[Lunderberg]

The GlobalVar instances themselves don't need to be changed, only the functions to which they point. This is why the pre-transform steps don't need to modify any function bodies, because the GlobalVar is still present in the IRModule. After running the transform, the values from new_subset are only copied back to the original mod if they weren't one of the dummy ExternFunc instances.

[Lunderberg]

The CI failure here will be resolved with #16843.

Edit: #16844 is the commit that will resolve the CI failure here.

[slyubomirsky]

Wait, a global var can be mapped to an ExternFunc? That's wild.

[slyubomirsky]

I had no idea that it was possible to have an ExternFunc as a module entry. If this is behavior used elsewhere, I should probably mention this in the specification. Otherwise, this "should" be safe because most accesses of global vars have to check if the value is a Relax Function since it could be a PrimFunc... then again, there might be hazards if a pass assumes (reasonable) that those are the only two possibilities. In any case, this is an improvement compared to passing a potentially malformed module altogether.

[Lunderberg]

Yup, relax::ExternFunc is a subclass of BaseFunc, and can be stored directly in an IRModule. At some point, I want to use this to specify when the lookup of a PackedFunc in the global registry occurs. Currently, there's a discrepency between Relax and TIR, where Relax performs the lookup when the VirtualMachine is initialized, while TIR performs the lookup just before the call is performed. I'd like to make these more similar, with an ExternFunc being resolved either at initialization (declared in IRModule), at the start of a function (ExternFunc bound to a variable), or at the callsite (ExternFunc used inline).

But the main impetus for that would have been the LazyTransformParams pass, and with the LazyGetInput accepting a callback parameter, there isn't as much need for it.

[Lunderberg]

Rebased this PR onto main, since #16844 has now landed.