[dynamo] Graph break on pack_padded_sequence (#108096)

This is to workaround #93501.

Fixes errors in:
```
./benchmarks/dynamo/torchbench.py --inference --performance --no-skip --inductor --freezing --only tacotron2
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/108096
Approved by: https://github.com/davidberard98

torch/_dynamo/variables/torch.py

@@ -655,6 +655,8 @@ class TorchVariable(VariableTracker):
                     return v
 
             return torch.utils._pytree.tree_map(map_fn, tree)
+        elif self.value is torch.nn.utils.rnn.pack_padded_sequence:
+            unimplemented("workaround https://github.com/pytorch/pytorch/issues/93501")
         elif isinstance(self.value, types.ModuleType):
             unimplemented("TypeError(\"'module' object is not callable\")")
         else:

torch/_inductor/select_algorithm.py

@@ -340,6 +340,7 @@ class TritonTemplateKernel(TritonKernel):
     def call_kernel(self, name: str):
         wrapper = V.graph.wrapper_code
         _, call_args, _ = self.args.python_argdefs()
+        call_args = [str(a) for a in call_args]
 
         for i in range(len(call_args)):
             if V.graph.is_unspec_arg(call_args[i]):
@@ -354,7 +355,7 @@ class TritonTemplateKernel(TritonKernel):
                 device_index=V.graph.scheduler.current_device.index,
             )
         else:
-            call_args = ", ".join(call_args)
+            call_args = ", ".join(call_args)  # type: ignore[assignment]
             stream_name = wrapper.write_get_cuda_stream(
                 V.graph.scheduler.current_device.index
             )