NaN-aware sum-of-squares (nanss)

[rhandberg]

Bottleneck already includes a fast sum-of-squares function (ss), but no NaN-aware counterpart exists. Wouldn't it make sense to add a nanss function?

[qwhelan]

Seems straightforward enough. I'm trying to pin down the exact casting behavior of the old scipy.stats.ss() (seeing some discrepancies with hypothesis fuzz testing), but once that's nailed down a nanss should be trivial.

Thanks for raising!