Disentangling Past-Future Modeling in Sequential Recommendation via Dual Networks