Improving predictive uncertainty estimation using Dropout–Hamiltonian Monte Carlo