fix: guard against division by zero in correlation()

[lippinj]

When both variances are near zero, sqrt(variance_x * variance_y) can underflow to zero and cause a division by zero. Clamp the divisor to epsilon instead.

[davisking]

Thanks @lippinj for identifying this underflow. I followed up with normalization that divides by the smaller standard deviation first and then the larger, avoiding the scale dependence of an epsilon clamp. I also documented the zero-variance result and added regression tests for tiny, large, and asymmetric scales.