Fix slicing list returning wrong result

[yanboliang]

This is the 3/N issue found from #156.

Problem:

• Failed test case: test_achaiah_pywick.py:RefineNet4Cascade
• Error: RuntimeError: Given groups=1, weight of size [256, 256, 3, 3], expected input[4, 512, 2, 2] to have 256 channels, but got 512 channels instead

Minimal code to reproduce:

from typing import List
import torch
import torchdynamo
import numpy as np

def fn(input):
    y = 0
    for i, x in enumerate(input[2:], 1):
        y = y + x
    return y

def my_compiler(gm: torch.fx.GraphModule, example_inputs: List[torch.Tensor]):
    return gm.forward

with torchdynamo.optimize(my_compiler):
    print(fn([1, 2, 3]))

The correct output of the above code is: 3.
However, it print out 6 in the current main.

Root cause and solution:

• If we print out the original bytecode and modified bytecode, we'll found BINARY_SUBSCR is missing. BINARY_SUBSCR is used to generate input[2:], so it outputs wrong result which is the sum of the whole list rather than the slice elements.
• Then dig into PyCodegen.__call__ and found the error comes from source reconstruction. The ListVariable.source should be GetItemSource but is LocalSource now. This list was created at BaseListVariable.getitem_const by cloning, but don't pass in the correct source.