Solving Constrained Combinatorial Optimisation Problems via MAP Inference without High-Order Penalties