Unify `src/mcmc/Inference.jl` methods

[penelopeysm]

There are several kinds of 'transitions' that can come out of samplers:

• Turing.Inference.Transition is the most common one.
• There are some other special transitions, SGLDTransition, SMCTransition, and PGTransition. These, together with Turing.Inference.Transition, get lumped under Turing.Inference.AbstractTransition. Fixed by Unify Turing Transitions, fix some tests #2651
• Some samplers also spit out varinfos (cough cough SampleFromPrior cough cough, and I think Prior also does this).

We should get rid of all of these conceptually similar things and just use a single Transition object.

I think this will have to wait for further changes in DynamicPPL because SampleFromPrior is not yet removed, but once that is done we can definitely get rid of the VarInfo methods.

The rest should also not be overly difficult to do.

[penelopeysm]

The joys of multiple dispatch... 🙄

Turing.jl/src/mcmc/Inference.jl

Lines 139 to 144 in 23b92eb

	Transition(θ, lp) = Transition(θ, lp, nothing)
	function Transition(model::DynamicPPL.Model, vi::AbstractVarInfo, t)
	θ = getparams(model, vi)
	lp = getlogp(vi)
	return Transition(θ, lp, getstats(t))
	end

If you called Transition(model, vi) you would get the method on line 139 even though that's clearly not what it was intended to do. I suppose that by coincidence it all works out...

[mhauru]

Would we rather solve this with TuringLang/AbstractMCMC.jl#149?

[penelopeysm]

Would be nice to, just don't know how close at hand that is.