Use more tfp distributions

In recent [updates](https://github.com/tensorflow/probability/releases) of tensorflow_probability the Pareto and LKJ distributions have been added. We currently include these distributions 'by hand'; should we rewrite the implementations of these using the new tfp distributions?

Also, I notice that the Cauchy distribution is also implemented 'by hand' despite it's inclusion in tfp (though there is probably a good reason for this)

All the best!