Add coupled multinomial HMC

[rlouf]

"Multinomial HMC" should be easy to implement once the NUTS PR is merged since we formulated NUTS as trajectory sampling.

https://arxiv.org/abs/2104.05134

We only need to implement a linear_expansion function that samples a direction and runs the integrator for L steps. Implement the particular case L=1 directly without scan.

See https://arxiv.org/abs/1701.02434 page 40. Choosing a sample from the entire trajectory supposedly is more efficient in terms of ESS/gradient evaluation.

[rlouf]

And thus rename the current version of HMC, "vintage_hmc"?

https://discourse.mc-stan.org/t/hmc-jittered-vs-nuts-on-1000-dimensional-standard-normal/8383/2