A Solver-Free Framework for Scalable Learning in Neural ILP Architectures