Unit tests on implemented metrics

[cwharland]

Nice looking package and it seems to bootstrap model comparisons in a nice way. You roll your own implementation of the metrics and while they are straightforward implementations the unit tests don't check for correctness at all. This makes it hard to adopt the package without first checking the correctness ourselves.

Seems like two paths could help improve:

1. Drop custom metrics in favor of already tested implementations (i.e. sklearn)
2. Add unit tests for current implementations

We'll try to carve out time to do that latter but not sure which direction you'd want to take.

Thanks for the work!