A cleaner and more scalable implementation of symbolic tracing (#11763)

Cleaner and more scalable implementation of symbolic tracing with torch.fx, and provides support for new architectures:
- ALBERT
- DistilBERT
- MobileBERT
- MegatronBERT
- GPT2
- GPT Neo

Co-authored-by: Michael Benayoun <michael@huggingface.co>

tests/test_modeling_mobilebert.py

@@ -267,6 +267,7 @@ class MobileBertModelTest(ModelTesterMixin, unittest.TestCase):
         if is_torch_available()
         else ()
     )
+    fx_ready_model_classes = all_model_classes
     test_sequence_classification_problem_types = True
 
     # special case for ForPreTraining model