Add support for pegasus model when using torch.compile(..., dynamic=True).

Fix some bugs:

1. The CompiledGraph does not support the graph that directly returns a constant tensor in fast path.
2. Move the correctness check to each run so that it also supports the checking for dynamic shape. Previously, we use dummy inputs and now we use the actual inputs. It is purely used for debugging the correctness of each individual operator.
3. Remove the FunctionBodyPass, as it will not update the function parameters when we update the tensor parameters.