Fix torch.compile FunctionalTensor inputs for higherOrderOps (#107604)
Before this PR, for the added [test](https://github.com/pytorch/pytorch/pull/107604/files#diff-c618f2274b6b5ccc533c580549d2e552edbd9fc5ac0da1aa4b00338525c8f78dR224), which feeds FunctionTensorWrapper inputs to higherOrderOperator, we have an assertion error in this line [code](https://github.com/pytorch/pytorch/pull/107604/files#diff-9f0663783bcd93e948e0491ef61b48123bdc9977bcc632fd707da578df13bfa1R1284).
The key difference of this PR is this [line ](https://github.com/pytorch/pytorch/pull/107604/files#diff-9f0663783bcd93e948e0491ef61b48123bdc9977bcc632fd707da578df13bfa1L1263)of check:
```python
elif (
isinstance(example_value, FakeTensor)
and example_value.fake_mode is tx.fake_mode
):
```
The original intention of it seems to be dealing with case where we want to wrap an fx proxy for an intermediate fake tensor that's produced by some tensor ops and an example value is provided (as is the case for higherOrderOps [here](https://github.com/pytorch/pytorch/blob/main/torch/_dynamo/variables/higher_order_ops.py#L85)). A fakified FunctionalTensorWrapper(FakeTensor) always fails this check. This PR changes it to checking whether it's already fakified by tx.fake_mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107604
Approved by: https://github.com/zou3519
ghstack dependencies: #107569
