Fixing question-answering with long contexts (#13873)

* Tmp.

* Fixing BC for question answering with long context.

* Capping model_max_length to avoid tf overflow.

* Bad workaround bugged roberta.

* Fixing name.

tests/test_modeling_reformer.py

@@ -189,6 +189,7 @@ class ReformerModelTester:
     def get_pipeline_config(self):
         config = self.get_config()
         config.vocab_size = 100
+        config.max_position_embeddings = 100
         config.axial_pos_shape = (4, 25)
         config.is_decoder = False
         return config