Fix expected test failures for predispatch export nested cond and out_dtype (#108715)
Before this PR, we use get_fake_value to get the fake_sub_args then call op(*fake_sub_args) to get the example value for out dtype.
This causes problem when the input proxy's op type is `get_attr`, get_fake_value for a `get_attr` node will actually look at the original param/buffer and **return a real tensor** instead of fake tensor. This is OK for export, since export's fake_mode allows non_fake_inputs see [here](https://github.com/pytorch/pytorch/blob/main/torch/_dynamo/output_graph.py#L278). But it causes problem when nesting cond with out_dtype where cond will use torch.compile(full_graph=True) to inspect out_dtype and find the inputs to op are mixed FakeTensor and real tensor.
This PR changes how we get the example values from proxies by directly looking at node.meta["example_value"]. This meta data is guaranteed to exist for all proxies during dynamo tracing so it's safe to use ( it's also used by get_fake_value to get fake tensors from args for general ops see [here](https://github.com/pytorch/pytorch/blob/main/torch/_dynamo/utils.py#L1318)).
Test Plan:
existing tests + remove expected failure for a test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108715
Approved by: https://github.com/zou3519
