A forward pass

Implement the forward pass of an MLP with any number of layers.

• relu(z): element-wise $\max (0,z)$.
• count_params(sizes): for layer sizes like [784, 128, 10], the total number of weights and biases.
• mlp_forward(X, params): params is a list of (W, b) pairs. Apply X @ W + b for each layer, with ReLU after every layer except the last. Return the final output (the logits).