Fix python gc race condition with THPVariable_traverse

[zou3519]

Fixes #3883: asserts being triggered in DataParallel for long training regiments like imagenet.

The bug is that a python Function in the computation graph gets garbage collected due to a C++ shared pointer to the C++ Function it wraps changing use_count() from 1 to more than 1 during python garbage collection. This happens because the "python state" of Variable changes: the THPVariable_traverse method indicates that when the shared pointer's use_count() is 1, the python Variable holds a python reference to the python Function, but not when use_count() > 1. See this gist for a more detailed exploration of the diagnosis.

The bug could be fixed by treating the shared pointer (a Variable's grad_fn) as part of the python state and intelligently grabbing the GIL when necessary. The downside to this approach is the complexity and potential overhead.

Alternatively, after talking with @colesbury, we think it might be better to remove the checking of the shared pointer's use_count() in THPVariable_traverse. This effectively disables reference cycle garbage collection for python Functions.

[apaszke]

Can you add a comment next to the deleted C++ code that links to your gist and maybe says a few words about why we don't want to traverse to the grad_fn. Also, mention that this in fact leaks when hooks are involved in a cycle