Refactor `output_hidden_states` to allow index selection

### Feature request Models such as Llava use CLIPVision with `output_hidden_states=True` then select the index of `hidden_states` that it needs e.g. `image_outputs.hidden_states[vision_feature_layer]`. `output_hidden_states` could be refactored to `output_hidden_states: Optional[Union[bool, int]] = None` and `hidden_states` output changed to `hidden_states: Optional[Union[torch.FloatTensor, Tuple[torch.FloatTensor, ...]]] = None`. The layer index would need to be normalized to account for negative index: ```python if type(output_hidden_states) is int and output_hidden_states < 0: output_hidden_states = min(len(self.layers), len(self.layers) + output_hidden_states + 1) # or config.num_hidden_layers ``` `CLIPEncoder` could be changed from ```python for idx, encoder_layer in enumerate(self.layers): if output_hidden_states: encoder_states = encoder_states + (hidden_states,) ``` to ```python for idx, encoder_layer in enumerate(self.layers): if type(output_hidden_states) is int and output_hidden_states == idx: encoder_states = hidden_states elif output_hidden_states: encoder_states = encoder_states + (hidden_states,) ``` and after the loop changed from ```python if output_hidden_states: encoder_states = encoder_states + (hidden_states,) ``` to ```python if type(output_hidden_states) is int and output_hidden_states == len(self.layers): encoder_states = hidden_states elif output_hidden_states: encoder_states = encoder_states + (hidden_states,) ``` Models such as Llava would then be able to do: ```python image_outputs = self.vision_tower(pixel_values, output_hidden_states=vision_feature_layer) selected_image_feature = image_outputs.hidden_states ``` ### Motivation Memory efficiency, as per the [comment](https://github.com/huggingface/transformers/blob/68049b17a6bb4c9b0d499e9e77121effa2f5a6c0/src/transformers/models/llava/modeling_llava.py#L454) in Llava `"this is not memory efficient at all (output_hidden_states=True) will …