Update constant fold to use correct numpy type (#2204)

In PyTorch<=2.7, the numpy arrays for bfloat16 and float8 types have
dtypes UINT16 and UINT8, which leads to incorrect constant folded
graphs. This PR updates the numpy helper to cast the arrays to the
correct dtypes.

Fix #2187

Author: justinchuby

onnxscript/optimizer/_constant_folding.py

@@ -297,7 +297,9 @@ def _get_numpy_value(
         if size_limit is not None and const_value.size > size_limit:
             return None
         try:
-            array = const_value.numpy()
+            # Reinterpret the array with `.view()` because some implementations of
+            # ir.TensorProtocol (e.g. PyTorch<=2.7) do not use ml_dtypes for bfloat16 etc.
+            array = const_value.numpy().view(const_value.dtype.numpy())
         except FileNotFoundError:
             # External data is not available.
             return None