Equilibrium Training of Energy-Based Models with Parallel Trajectory Tempering