Offline RL via Feature-Occupancy Gradient Ascent