BetaPrime `ln_pp` should depend on data

[cscherrer]

Currently in betaprime.rs we have

    fn ln_pp_cache(
        &self,
        data: &DataOrSuffStat<usize, StickBreakingDiscrete>,
    ) -> Self::PpCache {
        let post = self.posterior(data);
        post.alpha / post.beta
    }

    fn ln_pp_with_cache(&self, cache: &Self::PpCache, _y: &usize) -> f64 {
        cache.ln()
    }

Since ln_pp_with_cache doesn't depend on y, this would mean the posterior predictive is a weighted counting measure on usize, which doesn't make sense. It should be a proper distribution, summing to one.

See #14