Effective control of two-dimensional Rayleigh--B\'enard convection: invariant multi-agent reinforcement learning is all you need