What's the Move? Hybrid Imitation Learning via Salient Points