Fix jit tracing namedtuple (#29477)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29477
When passing in a namedtuple as trcing input, __clone_inputs will call into `torch.autograd.function._nested_map` and https://github.com/pytorch/pytorch/blob/593bb14/torch/autograd/function.py#L256 will run into error (because namedtuple doesn't support this style of constructor).
ghstack-source-id: 93586773
Differential Revision: D18405504
fbshipit-source-id: 8d0135cff0bdaaabcf6e06fac63df0f75c0c50b9
