Clarify documentation of labels_train/test usage wrt loss_train/test.

PiperOrigin-RevId: 665469121

tensorflow_privacy/privacy/privacy_tests/membership_inference_attack/data_structures.py

@@ -261,7 +261,9 @@ class AttackInputData:
   # Contains ground-truth classes. For single-label classification, classes are
   # assumed to be integers starting from 0. For multi-label classification,
   # label is assumed to be multi-hot, i.e., labels is a binary array of shape
-  # (num_examples, num_classes).
+  # (num_examples, num_classes). Simplifies testing for classification models
+  # by computing the loss_train/test internally. Not used if loss_train/test is
+  # provided.
   labels_train: Optional[np.ndarray] = None
   labels_test: Optional[np.ndarray] = None