support training deep dla model with relative position presentation by setting attention_type = "relative_dot_product"
