The old approach of getting state tries involved instantiating a
new state database and using it to instantiate a trie. What we didn't
realize was that state database needed to have Close() called on it,
which we didn't offer a way for plugins to do, resulting in memory
leaking.
This approach reuses the primary trie database associated with the
blockchain object, albeit a bit indirectly. This will allow access
to recent tries that are stored in memory, where previously only
tries that had been committed to disk were accessible.
When the plugin loader itself had to know the types in the arguments
and return values of the plugin functions, it was very difficult to
avoid import loops, given that the types were often defined in the
same package that needed to invoke the plugins.
Under this model, the plugin loader has much less knowledge of the
plugins themselves, and within each package we define functions to
interact with the plugins.
