Allowing small negative Eigen values in prior matrices

[nadrino]

For large matrices (4000x4000), decomposing often leads to a bunch of Eigen values close to zero and even slightly negative. These are due to numerical noise. The current version of GUNDAM stops the runtime asking to fix the matrix or apply a threshold on the Eigen values. This PR replaces it with an ALERT message letting numerical noise as it is

[ClarkMcGrew]

This isn't a good idea since we already have options to handle this. It's also super bad to "hide" this from analyzers. If they are using a bad covariance (almost degenerate), then they really need to control how they work around it, and we do provide the tools.

[nadrino]

I see, don't you think only keeping positive Eigen noise value would not introduce a bias somehow?

[ClarkMcGrew]

I'm not sure what it "means" for a covariance to have a negative eigenvalue. I think it's completely non-physical because it implies that a variable has a variance that isn't equal to it's own variance (yes, that's contradictory).

I think negative eigenvalues happen because of imprecision in the covariance, and should be interpreted as zero (and, if the value becomes too negative, there are real problems with the inputs).