Optimizing What Policies Learn From: Recoverability-aware Rollout Intervention Learning