Trajectory-Level Redirection Attacks on Vision-Language-Action Models