Raise clear error for problem_type="single_label_classification" with num_labels=1 (huggingface#45611)

* Raise clear error for problem_type="single_label_classification" with num_labels=1

This combination is mathematically degenerate: applying cross-entropy loss to a
single logit always yields zero loss, so training silently accomplishes nothing.
Validate the combination in PreTrainedConfig.__post_init__ so users get a clear
error at config construction with a pointer to the correct setup (num_labels=2
for binary classification, or problem_type="regression" for a single-output
regression head).

Closes huggingface#45479

* Update src/transformers/configuration_utils.py

* Update tests/utils/test_configuration_utils.py

* Update src/transformers/configuration_utils.py

---------

Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

Author: gaurav0107

src/transformers/configuration_utils.py

@@ -265,6 +265,13 @@ def __post_init__(self, **kwargs):
             # Keys are always strings in JSON so convert ids to int
             self.id2label = {int(key): value for key, value in self.id2label.items()}
 
+        if self.problem_type == "single_label_classification" and self.num_labels == 1:
+            raise ValueError(
+                '`problem_type="single_label_classification"` requires `num_labels > 1`. For binary '
+                'classification use `num_labels=2`, or use `problem_type="regression"` for a '
+                "single-output regression head."
+            )
+
         # BC for rotary embeddings. We will pop out legacy keys from kwargs and rename to new format
         if hasattr(self, "rope_parameters"):
             kwargs = self.convert_rope_params_to_dict(**kwargs)