Covariance-based decoding reveals content-specific feature integration and top-down processing during visual imagery