Big TF test cleanup (#24282)

* Fix one BLIP arg not being optional, remove misspelled arg

* Remove the lxmert test overrides and just use the base test_saved_model_creation

* saved_model_creation fixes and re-enabling tests across the board

* Remove unnecessary skip

* Stop caching sinusoidal embeddings in speech_to_text

* Fix transfo_xl compilation

* Fix transfo_xl compilation

* Fix the conditionals in xglm

* Set the save spec only when building

* Clarify comment

* Move comment correctly

* Correct embeddings generation for speech2text

* Mark RAG generation tests as @slow

* Remove redundant else:

* Add comment to clarify the save_spec line in build()

* Fix size tests for XGLM at last!

* make fixup

* Remove one band_part operation

* Mark test_keras_fit as @slow

tests/models/regnet/test_modeling_tf_regnet.py

@@ -148,6 +148,7 @@ class TFRegNetModelTest(TFModelTesterMixin, PipelineTesterMixin, unittest.TestCa
         not is_tf_available() or len(tf.config.list_physical_devices("GPU")) == 0,
         reason="TF does not support backprop for grouped convolutions on CPU.",
     )
+    @slow
     def test_keras_fit(self):
         super().test_keras_fit()