Skip nonzero unbacked SymInt memo in inference mode (#122147)
Summary: In `torch.inference_mode()`, fake tensors don't have `_version`s. This breaks unbacked SymInt memoization in `torch.nonzero` tracing. Here we disable the latter in inference mode.
Fixes https://github.com/pytorch/pytorch/issues/122127
Test Plan:
```
$ python test/inductor/test_unbacked_symints.py -k test_nonzero_in_inference_mode
...
----------------------------------------------------------------------
Ran 2 tests in 14.060s
OK
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122147
Approved by: https://github.com/ezyang
