BetaZero: Belief-State Planning for Long-Horizon POMDPs using Learned Approximations