Text this: Recursive GNNs for Learning Precoding Policies With Size-Generalizability