Unified Attention Modeling for Efficient Free-Viewing and Visual Search via Shared Representations