Does Hamiltonian Monte Carlo mix faster than a random walk on multimodal densities?