Generalised HMC with neural networks

[ben18785]

Another paper by Hoffman (inventor of NUTS), that generalises the leapfrog operator that allows them to be scaled dynamically.

[MichaelClerx]

It might be worth thinking about at some point whether we can generalise things like the leapfrog method to have some sort of re-usable leapfrog code that different algos can rely on

[MichaelClerx]

Then again, it's possible they all take different input arguments, produce slightly different output, and would make reading the code an absolute nightmare :D