Fix UnboundLocalError in RT-DETR loss computation (#42224)

* Fix UnboundLocalError in RT-DETR loss computation

Initialize auxiliary_outputs to None before conditional use to prevent
UnboundLocalError when config.auxiliary_loss is False.

Fixes the error:
  UnboundLocalError: local variable 'auxiliary_outputs' referenced before assignment

This occurs when auxiliary_loss is disabled but the variable is still
referenced later in the function.

* Update src/transformers/loss/loss_rt_detr.py

---------

Co-authored-by: Matt <[email protected]>

Author: yashwantbezawada

src/transformers/loss/loss_rt_detr.py

@@ -447,6 +447,7 @@ def RTDetrForObjectDetectionLoss(
     outputs_loss = {}
     outputs_loss["logits"] = logits
     outputs_loss["pred_boxes"] = pred_boxes
+    auxiliary_outputs = None
     if config.auxiliary_loss:
         if denoising_meta_values is not None:
             dn_out_coord, outputs_coord = torch.split(outputs_coord, denoising_meta_values["dn_num_split"], dim=2)