Merge pull request RasaHQ#5710 from RasaHQ/revert-5626-bug-activation

Revert "Move activation to transformer block"

Author: erohmensing

rasa/nlu/classifiers/diet_classifier.py

@@ -1250,10 +1250,7 @@ def _create_sequence(
         outputs = self._tf_layers[f"{name}_transformer"](
             transformer_inputs, 1 - mask, self._training
         )
-
-        if self.config[NUM_TRANSFORMER_LAYERS] > 0:
-            # apply activation
-            outputs = tfa.activations.gelu(outputs)
+        outputs = tfa.activations.gelu(outputs)
 
         return outputs, inputs, seq_ids, lm_mask_bool