Libcst code generation

[justinchuby]

• Remove assignment and return the last expression

   @script()
   def aten_split(self, split_size):
       val_return_val = op.SplitToSequence(val_self, val_split_size, axis=0)
       return return_val

  should become

   @script()
   def aten_split(self, split_size):
       return op.SplitToSequence(val_self, val_split_size, axis=0)

• Transform if graphs into branches

• Normalize invalid node names

• Handle function domain names: Currently the domain name is appended onto the function name, but that is not a valid name

• Use value_ints etc. to replace make_tensor calls

   val_414 = op.Constant(
       value=onnx.helper.make_tensor(
           name="value", data_type=onnx.TensorProto.INT64, dims=[2], vals=[-1, 8]
       )
   )

[BowenBao]

Let's consider this too #806. And should we inline small tensors, like the one in your example for Use value_ints etc. to replace make_tensor calls?

[gramalingam]

• Remove assignment and return the last expression

Since this is fairly generic, I wonder if we can just use some standard python tools as a post-processing cleanup (rather than implement it ourselves)?