Handle more variations in input graph structure

[caufieldjh]

Ontologies are special and each of them is different.
As per @cmungall in #4 :

avoid making any assumptions about root nodes.

Some false assumptions about ontologies:

• there is a single root node (unless owl:Thing is considered)
• there is a single graph
• the graph of all predicate types is a DAG
• the is-a graph is a DAG (subClassOf is reflexive, A=B <=> A sub B, B sub A
  etc

This means some ontologies, when used as input for semsim, may either be missing expected pairwise comparisons or the resulting scores may not reflect all manners of similarity defined by the ontology.
Examples abound with CHEBI (see chebi#is_conjugate_acid_of and chebi#is_conjugate_base_of relationships vs. is-a relationships).

[caufieldjh]

Also related to #8