Initial counter_cache implementation

[ekampp]

This implements a counter_cache option, which is passed to the parent relationship. This allows closure tree structures to use tag.tags_count rather than tag.children.count. This is much faster for large data sets, where counts are prohibitively slow.

TODO

• [ ] Sufficient performance
• [ ] Sufficient tests
• [ ] Write documentation for the feature

I would love the maintainers comments on the three TODO items above, to make sure that this PR doesn't break anything, and brings good enough performance and documentation.

Thanks, Emil

[ekampp]

This sort of picks up where https://github.com/ClosureTree/closure_tree/pull/109 left off. But it's re-implemented from scratch.