TF: BART compatible with XLA generation (#17479)

* Also propagate changes to blenderbot, blenderbot_small, marian, mbart, and pegasus

tests/models/gpt2/test_modeling_tf_gpt2.py

@@ -295,7 +295,7 @@ class TFGPT2ModelTester:
         self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
 
     def create_and_check_gpt2_xla_generate_fast(self, config, input_ids, *args):
-        config.eos_token_id = None
+        config.eos_token_id = None  # Generate until max length
         config.max_length = 10
         model = TFGPT2LMHeadModel(config=config)