Unify Trajectory type for static, dynamic HMC samplers.

[yebai]

We have the following nice composable interface for NUTS at the moment

p = NUTS{Multinomial,GeneralisedNoUTurn}(int)

where NUTS is a subtype of DynamicTrajectory. Maybe we can also unify StaticTrajectory and DynamicTrajectory via introducing a Static type and a MetropolisTS trajectory sampler (with the final point as an MH proposal), i.e.

struct Static{T<:Integer} <: AbstractTerminationCriterion
    n_leapfrog::T
end

struct MetropolisTS <: AbstractTrajectorySampler end

# dynamic HMC with Multinomial sampling and generalised No-U-Turn. 
nuts = Trajectory{MultinomialTS,GeneralisedNoUTurn}(int)

# static HMC with a Multinomial sampling
hmc_Multinomial = Trajectory{MultinomialTS,Static}(int)

# standard HMC with an MH correction step
hmc= Trajectory{MetropolisTS,Static}(int)

We can also use integration time instead of a number of leapfrog steps in the Static type. This allows us to use Trajectory to construct all common HMC samplers including HMC, HMCDA and NUTS.

[xukai92]

I will do #102 first and see how to unify them later.

[xukai92]

After working on the spiral idea for a bit I realise that what we really need to do is treat the transition kernel as a seperate type.

sturct Kernel{T<:AbstractTrajectory,S<:AbstractTrajectorySampler,F<:AbstractRefreshment}
t::T
s::S
f::F
end

The current NUTS{MultinomialTS,GeneralisedNoUTurn} should be Kernel(NUTS{GeneralisedNoUTurn}, MultinomialTS()) where f by default is set to be the completely refreshment.

By explicitly dispatch over transition() over Kernel, we can also the compelte control of a complete transition.

PS: We can also provide a composite kernel somehow. This is useful to combine a multinomial step + a MH step. Or maybe it's overcomplicated and we could introduce a specific type for this.

[yebai]

Closed by #245