Parallelize data augmentation

[osorensen]

The data augmentation step in Metropolis-Hastings is perfectly parallelizable. We can hence parallelize these two loops:

BayesMallows/src/augmentation_class.cpp

Lines 40 to 65 in db8a31b

	for(size_t i = 0; i < dat.n_assessors; ++i) {
	vec proposal;
	int g_diff{};
	if(pars.error_model == "none"){
	proposal = propose_pairwise_augmentation(
	dat.rankings.col(i), dat.items_above[i], dat.items_below[i]);
	} else if(pars.error_model == "bernoulli"){
	proposal = propose_swap(dat.rankings.col(i), dat.items_above[i],
	dat.items_below[i], g_diff, swap_leap);
	} else {
	Rcpp::stop("error_model must be 'none' or 'bernoulli'");
	}
	double u = std::log(R::runif(0, 1));
	int cluster = clus.current_cluster_assignment(i);
	double newdist = distfun->d(proposal, pars.rho_old.col(cluster));
	double olddist = distfun->d(dat.rankings.col(i), pars.rho_old.col(cluster));
	double ratio = -pars.alpha_old(cluster) / dat.n_items * (newdist - olddist);
	if(pars.error_model != "none") {
	ratio += g_diff * std::log(pars.theta(t) / (1 - pars.theta(t)));
	}
	if(ratio > u) dat.rankings.col(i) = proposal;
	}

and

BayesMallows/src/augmentation_class.cpp

Lines 75 to 83 in db8a31b

	for(size_t i = 0; i < dat.n_assessors; ++i){
	int cluster = clus.current_cluster_assignment(i);
	dat.rankings.col(i) = make_new_augmentation(
	dat.rankings.col(i), missing_indicator.col(i),
	pars.alpha_old(cluster), pars.rho_old.col(cluster),
	distfun, pseudo_aug_distance,
	log_aug_prob(i)
	);
	}

This depends on completing #366.