Learning to Abstract Visuomotor Mappings using Meta-Reinforcement Learning