Fix version counter sharing in set_data() (#20391)
Summary:
In https://github.com/pytorch/pytorch/pull/18223/files#diff-77a6f3462f2233b921d3042412fed6d3R178, we used `auto saved_version_ = data_.unsafeGetTensorImpl()->version_counter().current_version()` and then `new_data_impl_copy->set_version_counter(saved_version_)`, which actually doesn't preserve the original semantics that `var.set_data(tensor)` should keep `var`'s version counter object intact. This PR fixes the bug and adds test to make sure it doesn't happen again.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20391
Differential Revision: D15323430
Pulled By: yf225
fbshipit-source-id: e3ba49b51ec8ccecd51c80cb182387f74cfd2b2b