effective_size of nodes with only self-loop edges is undefined

[rossbar]

Handles the corner-case of isolated nodes that contain self-edges. Currently, isolated nodes are treated as having undefined effective_size:

>>> G = nx.Graph()
>>> G.add_node(0)
>>> nx.effective_size(G)
{0: nan}

However, the current behavior for isolated nodes with self-loops is inconsistent, resulting in ZeroDivisionError for the undirected case and 0.0 for directed graphs::

>>> G = nx.Graph([(0, 0)])
>>> nx.effective_size(G)
Traceback (most recent call last)
   ...
ZeroDivisionError: division by zero
>>> G = nx.DiGraph([(0, 0)])
>>> nx.effective_size(G)
{0: 0.0}

This discrepancy is due to the ambiguity of the direction of a self-edge.

Based on the current behavior and Borgatti's formula, I think the thing that makes the most sense is to have effective_size be undefined for all of these cases, making the behavior consistent across the board.

Closes #6916