Trajectory Data Suffices for Statistically Efficient Learning in Offline RL with Linear q \pi -Realizability and Concentrability