keep parameter names from PyTorch

[t-vi]

This patch uses PyTorch parameter names as the name hint in variables.
This means that one can load a stored PyTorch state dict and map the required inputs from that even after conversion.

[masahi]

Thanks @t-vi

[siju-samuel]

@t-vi Thanks for the PR.
I think this PR has slight impact in some scenarios.
You can confirm by running the below script before and after your change.

pytorch_pretrained_bert_uncased.py

I think after this PR, the param_tensor and param is not getting assinged properly in some cases..
Please look into this.
Thanks in advance.

[t-vi]

What is the error you are seeing?

[siju-samuel]

Output mismatch between pytorch and tvm.

Afer this PR

  File "pytorch_pretrained_bert_uncased.py", line 112, in <module>
    tvm.testing.assert_allclose(torch_preds, compiled_output, rtol=1e-3, atol=1e-3)
  File "/home/siju/workspace/tvm/python/tvm/testing.py", line 36, in assert_allclose
    np.testing.assert_allclose(actual, desired, rtol=rtol, atol=atol, verbose=True)
  File "/home/siju/.local/lib/python3.8/site-packages/numpy/testing/_private/utils.py", line 1532, in assert_allclose
    assert_array_compare(compare, actual, desired, err_msg=str(err_msg),
  File "/home/siju/.local/lib/python3.8/site-packages/numpy/testing/_private/utils.py", line 846, in assert_array_compare
    raise AssertionError(msg)
AssertionError: 
Not equal to tolerance rtol=0.001, atol=0.001

Mismatched elements: 427297 / 427308 (100%)
Max absolute difference: 24.692068
Max relative difference: 196537.89
 x: array([[[ -7.879808,  -7.787371,  -7.786093, ...,  -7.043789,
          -6.745376,  -4.60134 ],
        [-13.363304, -13.769426, -13.781861, ..., -11.81282 ,...
 y: array([[[-0.419126, -0.420205, -0.41907 , ..., -0.789973, -0.782199,
         -0.496477],
        [-0.419126, -0.420205, -0.41907 , ..., -0.789973, -0.782199,...

[t-vi]

I found the problem and will send a fix.
The old code included the embedding weight twice (which arguably is a bug in itself and needs fixing) and the new code deduplicated the param but not the var (which is even worse).
We underappreciate the structure of the PyTorch input...

[t-vi]

#5897 has the fix.