Add InverseWishart distribution

[jessegrabowski]

Description

This is an oft-requested distribution, and devs have always said no because it's not useful at all in the context of MCMC sampling (see here, also #4606, #975, #538) Our Wishhart distribution even warns you not to use it if you try!.

It seems, however, that the fine folks at Stan have cooked up a version that's parameterized with a cholesky decomposed input, which can be sampled using an appropriate transformation for lower-triangular matrices. Here are some comments about it by Bob Carpenter on the discourse.

I guess this would be analogous to our WishartBarlett distribution, which I honestly didn't know existed. That particular distribution should probably be introduced in rewrites, but it's out of scope for this issue.

One nice thing about having an InverseWishart distribution (as opposed to LKJ) is that we could rewrite Mvn(mu, IW) to PrecisionMvNormal and get some computational savings (avoid a matrix inverse). That seems nice. It also would have some implications for the experimental conjugate step samplers introduced in pymc-devs/pymc-extras#396