Amortized Bayesian Meta-Learning for Low-Rank Adaptation of Large Language Models