Scalability of Metropolis-within-Gibbs schemes for high-dimensional Bayesian models