Fix department label inconsistency (categorical vs sparse)

[koushalsmodi]

Title: Fix department label inconsistency in custom model example

Description: In one of the custom model examples, the department classification head is defined as:

self.department_classifier = layers.Dense(num_departments, activation="softmax")

with the model compiled using:

loss=["mean_squared_error", "categorical_crossentropy"]

This setup assumes one-hot encoded labels. However, the notebook currently generates labels using:

department_data = np.random.randint(0, num_departments, size=(num_samples, 1))

which produces integer indices, inconsistent with categorical_crossentropy.

---

Fix implemented in this PR:

• Replaced integer label generation with proper one-hot encoding using keras.utils.to_categorical.
• Added a comment showing that if integer labels are preferred, the correct loss is sparse_categorical_crossentropy.

---

Why this matters:

• Keeps the model output, labels, and loss function consistent.
• Prevents shape errors or misleading results.
• Improves clarity for readers following the example.

---

Updated code:

# Option 1: one-hot encoded labels (for categorical_crossentropy)
department_data = keras.utils.to_categorical(
    np.random.randint(0, num_departments, size=(num_samples,)), 
    num_classes=num_departments
)

# Option 2 (alternative): if you prefer integer labels, 
# then switch the model compile loss to "sparse_categorical_crossentropy"
# department_data = np.random.randint(0, num_departments, size=(num_samples, 1))