fix registration design for inductor graph partition for vLLM #165341
Description
Current design not working too well. For example, we think Inductor doesn't guarantee a way to get FakeTensors for an IR node, but we rely on needing that.
In PyTorch 2.9 we're going to monkeypatch PyTorch to work around this. We should use our learnings to evolve the design for PyTorch 2.10.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben