Efficient Open Domain Question Answering With Delayed Attention in Transformer-Based Models