Build a NumPy VGG-style block with two valid three-by-three convolutions, ReLU activations, and max pooling.
