Database layer refactor

[jsstevenson]

A few varied but related issues to wade through

• use singleton constructor for graph
• make metakb.database.Graph injectable in places like the QueryHandler
• feasibility of adhering to yield/close pattern for all cases where graph is used
• separate the graph access class and more specific data loading methods (i.e. those used by transformers)
• think about where we want transactional queries vs atomic queries

Generally, we want to focus on clean access to the graph, and to be very careful about any case where resources are manually managed and not automatically deconstructed. It's a little sloppy but improved in #356, but we can make it cleaner and more consistent with a refactor