Can trigger segfaults from Python IR interface

[ezyang]

As a general rule, it should not be possible to trigger a segfault from Python (unless you're using 'unsafe' APIs). The IR Python bindings do not uphold this property.

(/home/ezyang/local/pytorch/env) [[email protected] ~/local/pytorch] python
Python 3.6.2 |Continuum Analytics, Inc.| (default, Jul 20 2017, 13:51:32) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch._C
>>> g = torch._C.Graph()
>>> n = g.create("Add")
>>> del g
>>> n
Segmentation fault (core dumped)

This particular problem is because inherit C++'s memory management model, where a Node is only live as long as the Graph is live. This is OK in C++, but unacceptable in Python.