Bias from cluster-specific time trajectories in cluster-randomized stepped wedge trials