Fix 291

[hv10]

This is an attempt to fix #291

The solution I came up with after adding a test for my specific issue is to avoid the division by zero by adding OnlineStats.$\epsilon$ to the offending lines.

Only downside:
The basic test has to be modified to accept a solution which is slightly wrong (by $\epsilon$) instead of an exact match.

Aside from this it came up green for any other tests.

[joshday]

This is probably fine, but why do NaNs occur when fitting the same vector? I might want this behavior to be opt-in if fitting the same vector is breaking assumptions.

[hv10]

When fitting the same vector twice the variance between them is $0.0$ so all variance can be explained with just repeating the original vector as the principal component (when starting from a freshly initialized CCIPCA).
This should not be an issue for datasets where we expect at least some level of measurement noise (leading to a natural spread in the variables - and therefore variance) even if we measure the same underlying value twice.
In my case though one of my timeseries is without any measurement noise and holds a value for an amount of timesteps - leading to this issue.
As the CCIPCA method has a parameter describing it's forgetfullness it can even happen after a bunch of observations have already been fitted if we observe the same value for long enough.

I think the current implementation makes no assumption over the underlying data that would make it disallowed to fit to repeated observations.

Also: I have not checked if the issue also occurs when only one variable is of variance $0$.

Alternatively one could think about only "applying" an update to the CCIPCA when it does not lead to a eigenvalue of $0.0$.