Add tp_descr_get_impl infrastructure for descriptor protocol (#182213)

Summary:
Add infrastructure for modeling CPython's descriptor protocol in Dynamo:

- base.py: Route __get__ calls to tp_descr_get_impl when the VT implements it, mirroring CPython's slot_tp_descr_get.
- utils.py: Add is_torch_class helper to identify torch-internal classes whose C-level descriptors should go through trace_rules instead of descriptor VTs.

Authored with Claude.

X-link: pytorch/pytorch#182213
Approved by: https://github.com/guilhermeleobas

Reviewed By: huydhn

Differential Revision: D104188384

fbshipit-source-id: 3f9441312d389ece2d77c31beec8a9b05c35a1db

Author: anijain2305

userbenchmark/dynamo/dynamobench/_dynamo/utils.py

@@ -4974,6 +4974,19 @@ def is_tensor_getset_descriptor(name: str) -> bool:
         return False
 
 
+def is_torch_class(cls: type) -> bool:
+    """Check if cls is defined in torch or a torch submodule.
+
+    Torch-internal C-level descriptors (e.g. torch.Tensor.unsqueeze_) need
+    to go through VariableTracker.build / trace_rules so they get the right
+    VT (TorchInGraphFunctionVariable), which has guards like
+    inplace-view-on-input-tensor detection. This helper identifies classes
+    whose descriptors should take that path instead of descriptor VTs.
+    """
+    module = getattr(cls, "__module__", None)
+    return module is not None and (module == "torch" or module.startswith("torch."))
+
+
 def is_torch_function_object(value: Any) -> bool:
     return hasattr(value, "__torch_function__")