Improve numerical stability

[mleseach]

The numerical instability of integral arrays are well known, especially when all coefficients of the source array are positive.

This can be a problem especially when using Float32.

This is due to the fact that the integral table has coefficients of different magnitudes and their subtraction lead to numerical instability.

However a simple workaround to mitigate these disparities is to add a negative bias to the source array. In particular using the mean as a bias works quite well most of the time.

Here is a pluto notebook to illustrate the improvement of the numerical stability:
https://gist.github.com/mle-seach/04bb528455c941776c6084185e905666

Do you think this feature has a place in this package?

If so i can open a PR, however im not sure if i should change the current IntegralArray implementation or create a new type, nor if I should put a default bias like the mean.

[timholy]

Sorry, I only just noticed this now, and that gist is gone. But I think I "get the gist" anyway, and it's not a crazy idea. It would not have to be a separate type, as you'd only be changing internals and preserving the existing behavior (except for roundoff error).