Support marginals with 0 variance in multivariate Gaussian

[jucor]

Numpy supports it: should we? @shakirm, please, what do you think: useful for GP, or rare enough that we don't ned to make it a priority?

[jucor]

I love Lapack ! v 3.2 solved it. Now all that's left is to wrap this pstrf and benchmark its speed against portf -- see if we can always use pstrf instead of potrf.

[jucor]

@shakirm, look at that techreport from Lapack guys!! This is beautiful....

[jucor]

And voila :) With lazy version, though, until pstrf is wrapped in Torch: use potrf for Cholesky, and if fails, fallback on torch.svd.