NUTS could be implemented more efficiently and allow for user defined trajectory statistics

[nsiccha]

@penelopeysm: The other reason (apart from the implementation potentially being subtly wrong) I've never used NUTS.jl myself is because, even though it introduces less overhead than AdvancedHMC's or DynamicHMC's implementations, I needed something slightly more general than what I had implemented there.

If we want to re-implement NUTS at some point, I'd like it if we could have an eye towards that added generality.