improve implementation of `predict(...; include_all)`

[penelopeysm]

#984 uses init!! to implement predict. However, the implementation of include_all=false seems a bit wasteful because it first constructs a chain using all parameters (including the ones we don't want) before then subsetting the chain. It seems more sensible to, inside the loop, filter the dictionary of varname => value pairs in each iteration so that those variables don't end up in the chain to begin with.

DynamicPPL.jl/ext/DynamicPPLMCMCChainsExt.jl

Lines 122 to 158 in 956ed54

	predictive_samples = map(iters) do (sample_idx, chain_idx)
	# Extract values from the chain
	values_dict = chain_sample_to_varname_dict(parameter_only_chain, sample_idx, chain_idx)
	# Resample any variables that are not present in `values_dict`
	_, varinfo = DynamicPPL.init!!(
	rng,
	model,
	varinfo,
	DynamicPPL.InitFromParams(values_dict, DynamicPPL.InitFromPrior()),
	)
	vals = DynamicPPL.values_as_in_model(model, false, varinfo)
	varname_vals = mapreduce(
	collect,
	vcat,
	map(AbstractPPL.varname_and_value_leaves, keys(vals), values(vals)),
	)
	return (varname_and_values=varname_vals, logp=DynamicPPL.getlogjoint(varinfo))
	end
	chain_result = reduce(
	MCMCChains.chainscat,
	[
	_predictive_samples_to_chains(predictive_samples[:, chain_idx]) for
	chain_idx in 1:size(predictive_samples, 2)
	],
	)
	parameter_names = if include_all
	MCMCChains.names(chain_result, :parameters)
	else
	filter(
	k -> !(k in MCMCChains.names(parameter_only_chain, :parameters)),
	names(chain_result, :parameters),
	)
	end
	return chain_result[parameter_names]
	end

Not making this change in #984 to avoid complicating matters.

[penelopeysm]

I realised that this is pretty much impossible with MCMCChains because MCMCChains stores parameters like x[1], x[2], ... -- whereas values_as_in_model will have Dict(@varname(x) => ...) -- so it is not possible to figure out which parameters to drop from the dict!

This can be done with FlexiChains, but that's another story.