Describir: Multi-task reinforcement learning in reproducing Kernel Hilbert spaces via cross-learning