Exact Conversion of In-Context Learning to Model Weights in Linearized-Attention Transformers