Add new CheesHMC sampler by Hoffman et al

[twiecki]

http://proceedings.mlr.press/v130/hoffman21a/hoffman21a.pdf

[rlouf]

So I had a quick look. It is a new warmup scheme for the most part and here's what would be necessary:

1. a jittered version of HMC, which is trivial to implement here;
2. Dual averaging;
3. Mass matrix adaptation;
4. Their tuning algorithm for the scale of the trajectory length, T, which requires the Adam optimizer.

(2) and (3) are part of the Stan warmup and I already have an implementation of Adam in MCX that I can just copy/paste. Will implement once we have a drop in replacement for NUTS + Stan warmup.