[v1.5.x] Handle fix_gamma in tensorrt subgraph conversion correctly

[KellenSunderland]

Description

the current nnvm->onnx->tensorrt conversion is ignoring the fix_gamma field for batch norm layer. This commit tries to solve this particular issue by providing a way to massage params map before convert nnvm to onnx.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• [ x ] Changes are complete (i.e. I finished coding on this PR)
• [ x ] All changes have test coverage:
  • the unit test failed before and succeeded after

 a: array([[[[4.9751253]]]], dtype=float32)
 b: array([[[[0.]]]], dtype=float32)
-------------------- >> begin captured stdout << ---------------------
[[[[4.9751253]]]]
[[[[0.]]]]
--------------------- >> end captured stdout << ----------------------
----------------------------------------------------------------------
Ran 1 test in 5.820s
FAILED (failures=1)

.
----------------------------------------------------------------------
Ran 1 test in 5.513s
OK

• [ x ] Code is well-documented:
• [ x ] To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• [ x ] Add "Preprocess" step in nnvm->onnx conversion where allow modifying param map

Comments

This is a bug fix for tensor rt subgraph conversion

[pengzhao-intel]

LGTM

[TaoLv]

Thank you @KellenSunderland. Merging now.