Add MVN low rank #3038

Description

The benefit of a low-rank MVN is that the determinant and inversion of the covariance matrix is much cheaper. I also believe that this could be useful for models of MVNs that are not fitting the covariance matrix well which could be due to small sample sizes or actual low rank covariances. See pytorch for more details https://github.com/pytorch/pytorch/blob/14e348b7ad1b3472812f2b077020d80deaf6a787/torch/distributions/lowrank_multivariate_normal.py#L72.