Fix a bug where reordering tf.Transpose is not done correctly

[scottcjt]

When reordering tf.Transpose (downwards) and the original transpose must be preserved, the insertion point is off-by-one.

As demonstrated in attached test, this scenario is not handled properly:

%0 = transpose(%input, %axis_t)
%1 = pad(%0, %axis_p)
return %0, %1

Will get operand #0 does not dominate this use error from IR verifier, indicating a value is used before its def.

cc: @huangkang-chn

[smit-hinsu]

Thanks for the fix along with a test!