Add values to diagonal to deal with non-positive definite phylogenies

[noamross]

When fitting a funky phylogeny, I got the following error:

pen <- MRFtools::mrf_penalty(tree)
>chol.default(vcv(object)) : 
> the leading minor of order 178 is not positive definite

This can be solved by modifying the variance-covariance matrix slightly, making https://github.com/eric-pedersen/MRFtools/blob/master/R/penalties.R#L222 into:

eps <- .Machine$double.eps
pen <- chol2inv(chol(vcv(object) + eps*diag(length(object$tip.label)))) 

The question is, should eps be default or an argument to mrf_penalty.phylo() and should be this implemented for other methods? Or is this too finicky, in which case maybe there should be a .matrix method that takes a covariance matrix so people can provide it themselves as needed?

Happy to PR if we agree on the best way to do this.