Track dangling node change in indexing math generation.

[shmsong]

🐛 Describe the bug

This issue tracks the code drift related to index reference replay pass, see #1846 (comment) for the original discussion.

A notable difference observed is that all concrete indexing replay results are now free of dangling nodes:
https://github.com/csarofeen/pytorch/blob/devel/torch/csrc/jit/codegen/cuda/lower_index_compute.cpp#L782-L786

Would need to look further into if the dangling node part of loop indexing can be removed or if there are still cases we need to make sure they are added.

Versions

N/A

[shmsong]

More detail see #1846 (comment)