Turn shape type inference strict mode to false in optimizer (#1472)

Fix #1443 

In converter/dort, tensors retains their shape and type from PyTorch
models, and it saves us some efforts to infer them all like we did in
torchscript. However, when it comes to symbolic shapes, we still need
ONNX shape type inference. Error is raised when the inferred shape and
type are different from the carried ones. This is rare, but it happens
when a corner case is revealed. For example, in #1443, PyTorch generates
2 outputs with size=0 when native_batch_norm is run with CUDA.

This PR turn off the strict mode in ONNX shape type inference to avoid
crash in optimizer.

onnxscript/optimizer/__init__.py

@@ -58,8 +58,12 @@ def optimize(
     for _ in range(num_iterations):
         if onnx_shape_inference:
             if model.ByteSize() < 1024 * 1024 * 1024 * 2:
+                # NOTE: strict mode is disabled because it crashes on the models
+                # that have different shapes inferred from the model carried shapes.
+                # The case can be found in:
+                # https://github.com/microsoft/onnxscript/issues/1443
                 model = onnx.shape_inference.infer_shapes(
-                    model, check_type=True, strict_mode=True, data_prop=True
+                    model, check_type=True, strict_mode=False, data_prop=True
                 )
             else:
                 logger.warning(