Constrained Meta-Reinforcement Learning for Adaptable Safety Guarantee with Differentiable Convex Programming