Enable changing the loss function by making the hard-coded `loss_fct` an attribute of `BertForTokenClassification`.

### Feature request In the method `transformers.models.bert.modeling_bert.BertForTokenClassification.forward`, the `loss_fct = CrossEntropyLoss()` is currently hard-coded. To change the loss function (e.g., to set class weights in `CrossEntropyLoss`), one must currently monkey-patch the model. By making `loss_fct` an attribute (e.g., `self.loss_fct`), users can simply replace it and use custom loss functions during training. ### Motivation The motivation behind this proposal stems from the need to change the loss function for fine-tuning a pre-trained BERT model for token classification, particularly when dealing with imbalanced classes. In my use case, I need to prioritize recall, as most tokens belong to the "other" class. To achieve this, I need to set custom weights in the `CrossEntropyLoss`, like this: ```python loss_fct = CrossEntropyLoss(weight=torch.tensor([0.1, 1.0, 1.0, 2.0, 2.0], device=self.device) ``` However, since the loss function is hard-coded inside the `forward` method, modifying it currently requires overriding the entire method just to change one line, as shown here: ```python @patch def forward( self: BertForTokenClassification, input_ids: Optional[torch.Tensor] = None, attention_mask: Optional[torch.Tensor] = None, token_type_ids: Optional[torch.Tensor] = None, position_ids: Optional[torch.Tensor] = None, head_mask: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.Tensor] = None, labels: Optional[torch.Tensor] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple[torch.Tensor], 'TokenClassifierOutput']: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`. """ return…