Within-problem Learning for Efficient Lower Bound Computation in Max-SAT Solving