[matnormal] Matnormal covs should be ported to tf.linalg.LinearOperators when possible

[mshvartsman]

A lot of the structure-exploiting linear algebra in brainiak.matnormal is basically the same as in tf.linalg.LinearOperator. I'm not sure if I didn't know about it or it didn't exist in its current form when I did the original implementations, but we can probably remove a lot of code if most brainiak.matnormal.covs are just thin shims around LinearOperators. This would also buy us any future improvements on the TF side (in particular, moving scaling from cubic to quadratic if and when someone ports https://arxiv.org/pdf/1809.11165.pdf to TF), and batch support for covs. I have a few other things on my plate before getting to this (for one, porting MN-SRM to TF2) but this would be a nice way for someone else to get their hands dirty with brainiak.matnormal.