Mitigating Covariate Shift in Imitation Learning via Offline Data With Partial Coverage