FR: implement Poisson Binomial distribution

Non-identically distributed population of binomial variables, AKA Poisson's Binomial distribution (wiki), could have pmf naively computed with weighting over all combined priors for independent binomial variables, but I wouldn't expect this to be accurate for large number of success probabilities that aren't close to 0.5.

There's a memoization approach in the paper that's reference 5 in the wikipedia article. And I'd be happy to impl Rv, Support, DiscreteDistribution, that unless you have other ideas. Let me know!

I also don't know of a way to efficiently implement ln_f without computing f first, do you have any suggestions regarding this or how far to compute the memo table.