Support more model outputs for BERT/ERNIE/RoBERTa (#2583)

* Support more model outputs for BERT.

* Make more task classed support more outputs.

* Convert list to tuple automatically in ModelOutput.

* Support more model outputs for ERNIE.

* Support more model outputs for Roberta.

* Add embedding output to all hidden states.

* Clean import in model_output.py

Author: guoshengCS

paddlenlp/transformers/__init__.py

@@ -120,4 +120,11 @@
 if is_faster_tokenizer_available():
     from .bert.faster_tokenizer import *
     from .ernie.faster_tokenizer import *
-    from .tinybert.faster_tokenizer import *
+    from .tinybert.faster_tokenizer import *
+
+# patches of paddle.nn.Transformer to get all hidden_states and attentions
+import paddle
+from .model_outputs import _transformer_encoder_layer_fwd, _transformer_encoder_fwd
+
+paddle.nn.TransformerEncoderLayer.forward = _transformer_encoder_layer_fwd
+paddle.nn.TransformerEncoder.forward = _transformer_encoder_fwd