On Consensus-Optimality Trade-offs in Collaborative Deep Learning