Ten community algorithms error on networks holding signs as negative weights

[jhollway]

A sweep of every exported function over fict_marvel found that ten community
detection algorithms abort when called directly on a signed network:

node_in_betweenness   Weights must be strictly positive
node_in_eigen         Negative weight in weight vector
node_in_fluid         Negative weight in weight vector
node_in_greedy        Weights must not be negative
node_in_infomap       Edge weights must not be negative
node_in_labels        Weights must not be negative
node_in_louvain       Weight vector must not be negative
node_in_optimal       Negative weights are not allowed in weight vector
node_in_walktrap      Weight vector must be non-negative
node_in_automorphic   Negative cycle detected while calculating shortest paths

node_in_community() no longer fails on these, because it now considers only
node_in_spinglass() where the network is signed (#25). Called directly,
however, each still aborts.

Why this is not settled by the treatment in #25

Neither branch of the signed treatment obviously applies:

• .to_positive() is for measures of cohesion or distance. A community is not
  a distance, and dropping the negative ties discards exactly the hostility
  that should push two nodes into different communities.
• .to_unsigned() reads each tie by its magnitude, which is right for a census
  but here would silently treat hostility as affinity.

node_in_spinglass() is the one algorithm that reads a sign as a sign, through
igraph's implementation = "neg" (Traag and Bruggeman 2009). That may be the
answer for the family: either route every signed call there, or say plainly in
each function that the algorithm cannot read signs and name the alternative.

Related to #25.