Learning from Symmetry: Meta-Reinforcement Learning with Symmetric Data and Language Instructions