Bayesian bilevel operator learning with low-rank adaptation for efficient uncertainty quantification of PDE inverse problems