Avoid re-evaluating model in `getparams`

[penelopeysm]

This method currently requires re-evaluating the model:

Turing.jl/src/mcmc/Inference.jl

Lines 345 to 361 in 3901096

	function getparams(model::DynamicPPL.Model, vi::DynamicPPL.VarInfo)
	# NOTE: In the past, `invlink(vi, model)` + `values_as(vi, OrderedDict)` was used.
	# Unfortunately, using `invlink` can cause issues in scenarios where the constraints
	# of the parameters change depending on the realizations. Hence we have to use
	# `values_as_in_model`, which re-runs the model and extracts the parameters
	# as they are seen in the model, i.e. in the constrained space. Moreover,
	# this means that the code below will work both of linked and invlinked `vi`.
	# Ref: https://github.com/TuringLang/Turing.jl/issues/2195
	# NOTE: We need to `deepcopy` here to avoid modifying the original `vi`.
	vals = DynamicPPL.values_as_in_model(model, true, deepcopy(vi))
	# Obtain an iterator over the flattened parameter names and values.
	iters = map(DynamicPPL.varname_and_value_leaves, keys(vals), values(vals))
	# Materialize the iterators and concatenate.
	return mapreduce(collect, vcat, iters)
	end

Once TuringLang/DynamicPPL.jl#908 is merged and released, we should change it such that that accumulator is always used during model evaluation and change this method to just extract its contents.

[penelopeysm]

Unfortunately, we now need to re-evaluate the model to get both the user-space values as well as the logp breakdown, so this is no longer possible.

On the bright side, #2630 makes sure that all of that is obtained in a single evaluation, so we aren't doing extra extra evaluations.

[penelopeysm]

Note that it should still in principle be possible to completely circumvent any evaluations at all, but this would be contingent on using a richer interface than LogDensityProblems.jl.

The issue with LDP is that it expects as input an AbstractVector (representing values in internal space) and as output a single float (representing logjoint in internal space).

Thus when we call AbstractMCMC.step with a DynamicPPL.LogDensityFunction, we only get back the new values (in internal space) and the corresponding logjoint (in internal space).

If there was a way to get samplers to work with more complicated interface and return richer results, we would not need to re-evaluate the model at a point which the sampler already calculated.