Spatial-Language Attention Policies for Efficient Robot Learning