Use nn_module_stack instead

Summary: I don't know why source_fn_stack is disappeared but i think relying on nn_module_stack is better Differential Revision: D85959415

Author: tugsbayasgalan

torchao/quantization/pt2e/qat_utils.py

@@ -888,7 +888,7 @@ def _fold_conv_bn_qat(m: GraphModule) -> GraphModule:
  and node.args[0].op == "get_attr"
  and node.args[1] == 1
  and torch.nn.modules.batchnorm.BatchNorm2d
- in [val[1] for val in node.meta["source_fn_stack"]]
+ in [val[1] for _, val in node.meta["nn_module_stack"].items()]
  ):
  m.graph.erase_node(node)