$k$NN Prompting: Beyond-Context Learning with Calibration-Free Nearest Neighbor Inference