Allow returned tensor impls in fallback (#80545)
Previously, we had disallowed any reused storages in fallback kernels because the correct aliasing relationships weren't set up correctly. However, we can support the case of a reused TensorImpl by just returning the input FakeTensor it corresponds to. Additionally, we have `inplace_view` as a tag which will prevent us from running fallback kernels for kernels that mutate metadata. (thanks @ezyang who originally suggested I do this).
Fix for https://github.com/pytorch/torchdynamo/issues/467
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80545
Approved by: https://github.com/ezyang