divisi2.network is unnecessarily slow

[rspeer]

Right now when finding the nodes we might want to put into a matrix from a semantic network, we prune the undirected version of the network using networkx to make sure concepts have a high enough degree, put those into a matrix, and prune again using SparseMatrix.squish() to make sure features have a high enough degree as well.

If we could instead represent what goes into the matrix as a bipartite undirected NetworkX graph in the first place, then we could do the pruning right the first time, and make AnalogySpace faster.

[rspeer]

The most common case has been fixed because conceptnet_matrix() now just loads a reasonably-sized pickle file.