Fix Empirical *parameter* event ndims.

A *scalar* Empirical has self.event_ndims=0 and is parameterized by a *vector* samples of scalars.
A *matrix* Empirical has self.event_ndims=1 and is parameterized by a *matrix* samples of vectors.
Therefore we should be advertising _params_event_ndims=dict(self.event_ndims+1).

PiperOrigin-RevId: 237457364

tensorflow_probability/python/distributions/empirical.py

@@ -156,7 +156,7 @@ def _param_shapes(sample_shape):
     return {'samples': tf.convert_to_tensor(value=sample_shape, dtype=tf.int32)}
 
   def _params_event_ndims(self):
-    return dict(samples=self._event_ndims)
+    return dict(samples=self._event_ndims + 1)
 
   @property
   def samples(self):