Classification task: improve noise generation for target variable

[tail-call]

Instead of adding a random floating point value, we should replace class $q$ with class $r$ with a specified probability $p$.

Specifically see this in cgtnnlib/training.py:

        for i, (inputs, labels) in enumerate(dataset.data.train_loader):
            if is_classification_task(dataset.learning_task):
                pass
                # XXX
                # labels = add_noise_to_labels_classification(
                #     labels=labels,
                #     generate_sample=noise_generator.next_sample,
                # )
            else:
                labels = add_noise_to_labels_regression(
                    labels=labels,
                    generate_sample=noise_generator.next_sample,
                ).to(torch.float32