[functorch] Fix map() operator behavior. (#91906)
3 fixes made to control_flow.map:
1. argument list won't accept torch.nn.Module anymore, only Tensors.
2. during tracing we call new_empty from the returned sample output
instead xs to correctly inherit tensor metadata.
3. for FakeTensorMode we implement map() using new_empty() as well
instead of torch.stack() to preserve symbolic shape output.
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91906
Approved by: https://github.com/tugsbayasgalan