Bayesian Meta-Learning for Few-Shot Policy Adaptation Across Robotic Platforms