Add gradient transformation wrapper to be used in canned estimators.

Example:
    optimizer = some_optimizer
    optimizer = tf.contrib.estimator.clip_gradients_by_norm(optimizer, norm)
    estimator = tf.estimator.DNNClassifier(
        feature_columns=[...],
        hidden_units=[1024, 512, 256],
        optimizer=optimizer)

PiperOrigin-RevId: 169477090