Constraint-Informed Active Learning for End-to-End ACOPF Optimization Proxies