Deep Network Approximation: Beyond ReLU to Diverse Activation Functions