Bias should be resized with the weights

Created a link between the linear layer bias and the model attribute bias. This does not change anything for the user nor for the conversion scripts, but allows the `resize_token_embeddings` method to resize the bias as well as the weights of the decoder.

Added a test.

tests/test_modeling_common.py

@@ -487,6 +487,8 @@ class ModelTesterMixin:
             self.assertEqual(model.config.vocab_size, model_vocab_size + 10)
             # Check that it actually resizes the embeddings matrix
             self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] + 10)
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            model(**inputs_dict)
 
             # Check that resizing the token embeddings with a smaller vocab size decreases the model's vocab size
             model_embed = model.resize_token_embeddings(model_vocab_size - 15)
@@ -494,6 +496,11 @@ class ModelTesterMixin:
             # Check that it actually resizes the embeddings matrix
             self.assertEqual(model_embed.weight.shape[0], cloned_embeddings.shape[0] - 15)
 
+            # Check that the model can still do a forward pass successfully (every parameter should be resized)
+            # Input ids should be clamped to the maximum size of the vocabulary
+            inputs_dict["input_ids"].clamp_(max=model_vocab_size - 15 - 1)
+            model(**inputs_dict)
+
             # Check that adding and removing tokens has not modified the first part of the embedding matrix.
             models_equal = True
             for p1, p2 in zip(cloned_embeddings, model_embed.weight):