Text this: RL Perceptron: Generalization Dynamics of Policy Learning in High Dimensions