Filter-Based Abstractions with Correctness Guarantees for Planning under Uncertainty