TFP: fix key reuse issue in monte_carlo_variational_loss

Detected with JAX's enable_key_reuse_checks. We can avoid falling afoul of the reuse checker by splitting only once we determine it to be necessary. This should not have any user-visible change.

PiperOrigin-RevId: 613666004

tensorflow_probability/python/vi/csiszar_divergence.py

@@ -1074,7 +1074,6 @@ def monte_carlo_variational_loss(
       raise TypeError('`target_log_prob_fn` must be a Python `callable`'
                       'function.')
 
-    sample_seed, target_seed = samplers.split_seed(seed, 2)
     reparameterization_types = tf.nest.flatten(
         surrogate_posterior.reparameterization_type)
     if gradient_estimator is None:
@@ -1089,6 +1088,8 @@ def monte_carlo_variational_loss(
                            num_draws=importance_sample_size,
                            num_batch_draws=sample_size,
                            seed=seed)
+
+    sample_seed, target_seed = samplers.split_seed(seed, 2)
     if gradient_estimator == GradientEstimators.SCORE_FUNCTION:
       if tf.get_static_value(importance_sample_size) != 1:
         # TODO(b/213378570): Support score function gradients for