correlation loses precision on offset low-variance columns

[skymanbp]

The one-pass nSxy - SxSy form cancels catastrophically:

a = [1e8, 1e8, 1.00000002e8]
b = [1e8, 1e8, 1.00000003e8]
correlation "a" "b" df  -- Just 2.0

a = [1e11, 1e11 + 1, 1e11 + 2]
correlation "a" "a" df  -- Just (-1.0)

A centered two-pass form fixes both but doubles the passes, so it needs
profiling first (#214 review). Will send a PR with benchmarks.

[daikonradish]

How often does this happen that it needs a fixing?

[skymanbp]

How often does this happen that it needs a fixing?

Any column whose spread is small next to its mean. The error in the onepass form grows like eps * (mean/std)^2: at mean/std of about 1e4 you've lost half the digits, at 1e8 (the repro) it's pure cancellation. That shape is common, where you got timestamps (epoch seconds ~1.7e9 varying by seconds), prices around a large level, sensor data with a DC offset. And it fails silently when you get a plausible-looking r, or |r| > 1. numpy's corrcoef goes thru cov, which centers, for the same reason