Fix: Override output layer initializer to glorot_normal for softmax compatibility

[MohamedMostafa259]

In the original code, the RegularizedDense wrapper applies he_normal initialization to all layers, including the output layer. However, he_normal is optimized for ReLU activations and is not ideal for the final softmax layer according to Table 11-1 (Initialization parameters for each type of activation function) on page 360.

This PR updates the last layer to use glorot_normal initialization, which is better suited for layers with softmax activation. The change is small but improves clarity and correctness by aligning the initialization strategy with the activation function used.

Summary of changes

Before:

RegularizedDense(10, activation="softmax")

After:

RegularizedDense(10, activation="softmax", kernel_initializer="glorot_normal")