Is model pruning a sensible (or even possible) second model optimization method after the model has already been quantized?