STDEV is numerically unstable

[senderista]

Apparently our simple method in DoubleAggregator etc. of calculating running standard deviation is numerically unstable: https://www.johndcook.com/blog/2008/09/26/comparing-three-methods-of-computing-standard-deviation/. There is another one-pass method that is well-behaved (although incomprehensible): https://www.johndcook.com/blog/standard_deviation/. From the experiments described in these blog posts, I expect our current implementation to blow up for lots of users.