Towards Effective Context for Meta-Reinforcement Learning: an Approach based on Contrastive Learning