Quantifying Consistency and Information Loss for Causal Abstraction Learning