graph: fix bug in parsing of _output_shapes attr

[nfelt]

This fixes an issue where the Graph dashboard would die with "Graph: Failed Normalizing names" when trying to parse a graphdef containing a node with a mis-formatted _output_shapes attr. In particular, it fails if the _output_shapes attr doesn't contain the top-level list property (either it's empty, or has only other properties). The fix is just to guard the logic by checking for that property's existence before dereferencing it.

This attr is generated when calling as_graph_def(add_shapes=True) and I'm pretty sure that when constructed normally it should always have the .list.shape property if it's present on the node at all:

• https://cs.opensource.google/tensorflow/tensorflow/+/master:tensorflow/python/framework/ops.py;l=3304;drc=5e0c8476f0d8473e6246897ad1065ec4621bf73e
• https://cs.opensource.google/tensorflow/tensorflow/+/master:tensorflow/python/framework/ops.py;l=3347;drc=5e0c8476f0d8473e6246897ad1065ec4621bf73e

However, it seems like occasionally users end up with an attr value where that isn't the case, and even in those cases we should avoid breaking the graph visualization.

Googlers, see b/160928542.

[psybuzz]

Thanks for the fix and adding a test!