Better error calculation on marginal statistics

[htjb]

Currently, we input samples from the trained MAF or KDE to calculate the marginal statistics and samples from the original posterior then compare the results to get an error. I think the code needs a bit of a tidy, and it would be good to think about whether there is a better way to get an error on these statistics e.g. some monte carlo error.

I also think that the option to put in prior samples and prior weights to marginal_stats.calculate().statistics() should be scrapped in favour of supplying a prior density estimator so that the user can access all the fun stuff like early stopping and clustering.