[ONNX] Do not run 'deduplicate_initializers' when 'keep_initializers_…

…as_inputs' is True (#96320)

### Proposal
When arg of 'keep_initializers_as_inputs' is True, it's quite possible that parameters are set by initializer of input.
Hence we should disable de-duplicate initializer optimization when 'keep_initializers_as_inputs==True'.

- [x] Update doc related to `keep_initializers_as_inputs`.
Pull Request resolved: #96320
Approved by: https://github.com/abock, https://github.com/thiagocrepaldi

torch/onnx/utils.py

@@ -454,6 +454,11 @@ def forward(self, x):
             This may allow for better optimizations (e.g. constant folding) by
             backends/runtimes.
 
+            If True, `deduplicate_initializers` pass will not be executed. This means
+            initializers with duplicated values will not be deduplicated and
+            will be treated as distinct inputs to the graph. This allows different
+            input initializers to be supplied at the runtime following export.
+
             If ``opset_version < 9``, initializers MUST be part of graph
             inputs and this argument will be ignored and the behavior will be
             equivalent to setting this argument to True.
@@ -1603,9 +1608,11 @@ def _export(
                     module_typenames_to_export_as_functions,
                     list(params_dict.keys()),
                 )
-            params_dict = _C._jit_pass_onnx_deduplicate_initializers(  # type: ignore[assignment]
-                graph, params_dict, getattr(model, "training", False)  # type: ignore[arg-type]
-            )
+
+            if keep_initializers_as_inputs is not True:
+                params_dict = _C._jit_pass_onnx_deduplicate_initializers(  # type: ignore[assignment]
+                    graph, params_dict, getattr(model, "training", False)  # type: ignore[arg-type]
+                )
             _C._jit_pass_onnx_assign_scoped_names_for_node_and_value(graph)
             if export_params:
                 (